Re: [PATCH v6 00/10] s390: vfio-ap: dynamic configuration support

[Tony Krowiak <akrowiak@linux.ibm.com>]

On 10/8/19 8:57 AM, Pierre Morel wrote:
> 
> On 10/8/19 12:48 PM, Halil Pasic wrote:
>> On Fri, 13 Sep 2019 17:26:48 -0400
>> Tony Krowiak <akrowiak@linux.ibm.com> wrote:
>>
>>> The current design for AP pass-through does not support making dynamic
>>> changes to the AP matrix of a running guest resulting in three 
>>> deficiencies
>>> this patch series is intended to mitigate:
>>>
>>> 1. Adapters, domains and control domains can not be added to or removed
>>>     from a running guest. In order to modify a guest's AP configuration,
>>>     the guest must be terminated; only then can AP resources be assigned
>>>     to or unassigned from the guest's matrix mdev. The new AP 
>>> configuration
>>>     becomes available to the guest when it is subsequently restarted.
>>>
>>> 2. The AP bus's /sys/bus/ap/apmask and /sys/bus/ap/aqmask interfaces can
>>>     be modified by a root user without any restrictions. A change to 
>>> either
>>>     mask can result in AP queue devices being unbound from the vfio_ap
>>>     device driver and bound to a zcrypt device driver even if a guest is
>>>     using the queues, thus giving the host access to the guest's private
>>>     crypto data and vice versa.
>>>
>>> 3. The APQNs derived from the Cartesian product of the APIDs of the
>>>     adapters and APQIs of the domains assigned to a matrix mdev must
>>>     reference an AP queue device bound to the vfio_ap device driver.
>>>
>>> This patch series introduces the following changes to the current design
>>> to alleviate the shortcomings described above as well as to implement 
>>> more
>>> of the AP architecture:
>>>
>>> 1. A root user will be prevented from making changes to the AP bus's
>>>     /sys/bus/ap/apmask or /sys/bus/ap/aqmask if the ownership of an APQN
>>>     changes from the vfio_ap device driver to a zcrypt driver when 
>>> the APQN
>>>     is assigned to a matrix mdev.
>>>
>>> 2. The sysfs bind/unbind interfaces will be disabled for the vfio_ap
>>>     device driver.
>>>
>> Doesn't this have the potential of breaking some userspace stuff that
>> might be out there?
>>
>>> 3. Allow AP resources to be assigned to or removed from a matrix mdev
>>>     while a guest is using it and hot plug the resource into or hot 
>>> unplug
>>>     the resource from the running guest.
>> This looks like a natural extension of the interface -- i.e. should not
>> break any userspace.
>>
>>> 4. Allow assignment of an AP adapter or domain to a matrix mdev even 
>>> if it
>>>     results in assignment of an APQN that does not reference an AP queue
>>>     device bound to the vfio_ap device driver, as long as the APQN is 
>>> owned
>>>     by the vfio_ap driver. Allowing over-provisioning of AP resources
>>>     better models the architecture which does not preclude assigning AP
>>>     resources that are not yet available in the system. If/when the 
>>> queue
>>>     becomes available to the host, it will immediately also become 
>>> available
>>>     to the guest.
>> Same here -- I don't think this change breaks any userspace.
>>
>>> 1. Rationale for changes to AP bus's apmask/aqmask interfaces:
>>> ----------------------------------------------------------
>>> Due to the extremely sensitive nature of cryptographic data, it is
>>> imperative that great care be taken to ensure that such data is secured.
>>> Allowing a root user, either inadvertently or maliciously, to configure
>>> these masks such that a queue is shared between the host and a guest is
>>> not only avoidable, it is advisable.
> 
> Just curious: how is it possible to do such a configuration?

In the current implementation of dedicated crypto, there is nothing
stopping a sysadmin from changing the apmask/aqmask in manner that
transfers ownership of one more APQNs from the vfio_ap device driver to
zcrypt which results in unbinding the queue devices from vfio_ap and
binding them to the zcrypt drive. If a guest happens to be using the
queue at the time, both the host and guest will have access. That is
fixed by this series.

> 
> 
>>>   It was suggested that this scenario
>>> is better handled in user space with management software, but that does
>>> not preclude a malicious administrator from using the sysfs interfaces
>>> to gain access to a guest's crypto data. It was also suggested that this
>>> scenario could be avoided by taking access to the adapter away from the
>>> guest and zeroing out the queues prior to the vfio_ap driver 
>>> releasing the
>>> device; however, stealing an adapter in use from a guest as a by-product
>>> of an operation is bad and will likely cause problems for the guest
>>> unnecessarily. It was decided that the most effective solution with the
>>> least number of negative side effects is to prevent the situation at the
>>> source.
> 
> 
> Stealing an adapter in use by a guest, insn't it what is done if we 
> allow to unassign an AP/Domain using the unassign sysfs interface when 
> the mediated device is in use by the guest?

Yes, but that is a deliberate action as opposed to a side effect of
bind/unbind. It is the very definition of dynamic configuration (a.k.a.,
hot plug/unplug).

> 
> 
>>>
>>> 2. Rationale for disabling bind/unbind interfaces for vfio_ap driver:
>>> -----------------------------------------------------------------
>>> By disabling the bind/unbind interfaces for the vfio_ap device driver,
>>> the user is forced to use the AP bus's apmask/aqmask interfaces to 
>>> control
>>> the probing and removing of AP queues. There are two primary reasons for
>>> disabling the bind/unbind interfaces for the vfio_ap device driver:
>>>
>>> * The device architecture does not provide a means to prevent unbinding
>>>    a device from a device driver, so an AP queue device can be unbound
>>>    from the vfio_ap driver even when the queue is in use by a guest. By
>>>    disabling the unbind interface, the user is forced to use the AP 
>>> bus's
>>>    apmask/aqmask interfaces which will prevent this.
>>>
>> Isn't this fixed by your filtering (if implemented correctly)? BTW I 
>> believe
>> it solves the problem regardless whether the unbind was triggered by the
>> drivers unbind attribute or by a[pq]mask
>>
>>> * Binding of AP queues is controlled by the AP bus /sys/bus/ap/apmask 
>>> and
>>>    /sys/bus/ap/aqmask interfaces. If the masks indicate that an APQN is
>>>    owned by zcrypt, trying to bind it to the vfio_ap device driver will
>>>    fail; therefore, the bind interface is somewhat redundant and 
>>> certainly
>>>    unnecessary.
>>>
>>>
>>> Tony Krowiak (10):
>>>    s390: vfio-ap: Refactor vfio_ap driver probe and remove callbacks
>>>    s390: vfio-ap: allow assignment of unavailable AP resources to mdev
>>>      device
>>>    s390: vfio-ap: allow hot plug/unplug of AP resources using mdev 
>>> device
>>>    s390: vfio-ap: filter CRYCB bits for unavailable queue devices
>>>    s390: vfio-ap: sysfs attribute to display the guest CRYCB
>>>    s390: vfio-ap: update guest CRYCB in vfio_ap probe and remove
>>>      callbacks
>>>    s390: zcrypt: driver callback to indicate resource in use
>>>    s390: vfio-ap: implement in-use callback for vfio_ap driver
>>>    s390: vfio-ap: added versioning to vfio_ap module
>>>    s390: vfio-ap: update documentation
>> I believe it would be worthwhile to reorder the patches (fixes and
>> re-factoring first, the features).
>>
>> Regards,
>> Halil
>>
>>>   Documentation/s390/vfio-ap.rst        | 899 
>>> +++++++++++++++++++++++++---------
>>>   drivers/s390/crypto/ap_bus.c          | 144 +++++-
>>>   drivers/s390/crypto/ap_bus.h          |   4 +
>>>   drivers/s390/crypto/vfio_ap_drv.c     |  27 +-
>>>   drivers/s390/crypto/vfio_ap_ops.c     | 610 ++++++++++++++---------
>>>   drivers/s390/crypto/vfio_ap_private.h |  12 +-
>>>   6 files changed, 1200 insertions(+), 496 deletions(-)
>>>