Date:   Wed, 22 Jun 2022 12:42:51 +0200
From:   Halil Pasic <pasic@...ux.ibm.com>
To:     Christian Borntraeger <borntraeger@...ibm.com>
Cc:     Tony Krowiak <akrowiak@...ux.ibm.com>, linux-s390@...r.kernel.org,
        linux-kernel@...r.kernel.org, kvm@...r.kernel.org,
        jjherne@...ux.ibm.com, freude@...ux.ibm.com, cohuck@...hat.com,
        mjrosato@...ux.ibm.com, alex.williamson@...hat.com,
        kwankhede@...dia.com, fiuczy@...ux.ibm.com,
        Halil Pasic <pasic@...ux.ibm.com>
Subject: Re: [PATCH v20 00/20] s390/vfio-ap: dynamic configuration support

On Wed, 22 Jun 2022 08:48:27 +0200
Christian Borntraeger <borntraeger@...ibm.com> wrote:

> Am 21.06.22 um 17:51 schrieb Tony Krowiak:
> > The current design for AP pass-through does not support making dynamic
> > changes to the AP matrix of a running guest resulting in a few
> > 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. The
> >      AP architecture allows assignment of AP resources that are not
> >      available to the system, so this artificial restriction is not
> >      compliant with the architecture.
> > 
> > 4. The AP configuration profile can be dynamically changed for the linux
> >      host after a KVM guest is started. For example, a new domain can be
> >      dynamically added to the configuration profile via the SE or an HMC
> >      connected to a DPM enabled lpar. Likewise, AP adapters can be
> >      dynamically configured (online state) and deconfigured (standby state)
> >      using the SE, an SCLP command or an HMC connected to a DPM enabled
> >      lpar. This can result in inadvertent sharing of AP queues between the
> >      guest and host.
> > 
> > 5. A root user can manually unbind an AP queue device representing a
> >      queue in use by a KVM guest via the vfio_ap device driver's sysfs
> >      unbind attribute. In this case, the guest will be using a queue that
> >      is not bound to the driver which violates the device model.
> > 
> > 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 edits to the AP bus's
> >      /sys/bus/ap/apmask or /sys/bus/ap/aqmask if the change would transfer
> >      ownership of an APQN from the vfio_ap device driver to a zcrypt driver
> >      while the APQN is assigned to a matrix mdev.
> > 
> > 2. Allow a root user to hot plug/unplug AP adapters, domains and control
> >      domains for a KVM guest using the matrix mdev via its sysfs
> >      assign/unassign attributes.
> > 
> > 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 not reserved for use by the default zcrypt drivers (also known as
> >      over-provisioning of AP resources). 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. Such
> >      APQNs, however, will not be assigned to the guest using the matrix
> >      mdev; only APQNs referencing AP queue devices bound to the vfio_ap
> >      device driver will actually get assigned to the guest.
> > 
> > 5. Handle dynamic changes to the AP device model.
> > 
> > 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. 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.
> > 
> > 2. Rationale for hot plug/unplug using matrix mdev sysfs interfaces:
> > ----------------------------------------------------------------
> > Allowing a user to hot plug/unplug AP resources using the matrix mdev
> > sysfs interfaces circumvents the need to terminate the guest in order to
> > modify its AP configuration. Allowing dynamic configuration makes
> > reconfiguring a guest's AP matrix much less disruptive.
> > 
> > 3. Rationale for allowing over-provisioning of AP resources:
> > -----------------------------------------------------------
> > Allowing assignment of AP resources to a matrix mdev and ultimately to a
> > guest better models the AP architecture. The architecture does not
> > preclude assignment of unavailable AP resources. If a queue subsequently
> > becomes available while a guest using the matrix mdev to which its APQN
> > is assigned, the guest will be given access to it. If an APQN
> > is dynamically unassigned from the underlying host system, it will
> > automatically become unavailable to the guest.
> > 
> > Change log v19-v20:
> > ------------------
> > * Fixed patch 02/20: failed to move creation of status attribute
> >    for a queue device to the vfio_ap_mdev_probe_queue function in
> >    drivers/s390/crypto/vfio_ap_ops.c. (Jason)
> > 
> > * Fixed signature of get_update_locks_for_queue macro
> > 
> > * Take lock in get_update_locks_for_queue macro before
> >    accessing q->matrix_mdev
> > 
> > * Renamed vfio_ap_mdev_get_update_locks_for_apqn function to
> >    get_update_locks_for_apqn (Jason)
> > 
> > * Fix comments in function implementing the AP bus's in_use callback (Jason)
> > 
> > * Fix function name in prologue for ap_owned_by_def_drv function
> > 
> > Tony Krowiak (20):
> >    s390/vfio-ap: use new AP bus interface to search for queue devices
> >    s390/vfio-ap: move probe and remove callbacks to vfio_ap_ops.c
> >    s390/vfio-ap: manage link between queue struct and matrix mdev
> >    s390/vfio-ap: introduce shadow APCB
> >    s390/vfio-ap: refresh guest's APCB by filtering AP resources assigned
> >      to mdev
> >    s390/vfio-ap: allow assignment of unavailable AP queues to mdev device
> >    s390/vfio-ap: rename matrix_dev->lock mutex to matrix_dev->mdevs_lock
> >    s390/vfio-ap: introduce new mutex to control access to the KVM pointer
> >    s390/vfio-ap: use proper locking order when setting/clearing KVM
> >      pointer
> >    s390/vfio-ap: prepare for dynamic update of guest's APCB on
> >      assign/unassign
> >    s390/vfio-ap: prepare for dynamic update of guest's APCB on queue
> >      probe/remove
> >    s390/vfio-ap: allow hot plug/unplug of AP devices when
> >      assigned/unassigned
> >    s390/vfio-ap: hot plug/unplug of AP devices when probed/removed
> >    s390/vfio-ap: reset queues after adapter/domain unassignment
> >    s390/vfio-ap: implement in-use callback for vfio_ap driver
> >    s390/vfio-ap: sysfs attribute to display the guest's matrix
> >    s390/vfio-ap: handle config changed and scan complete notification
> >    s390/vfio-ap: update docs to include dynamic config support
> >    s390/Docs: new doc describing lock usage by the vfio_ap device driver
> >    MAINTAINERS: pick up all vfio_ap docs for VFIO AP maintainers
> > 
> >   Documentation/s390/vfio-ap-locking.rst |  105 ++
> >   Documentation/s390/vfio-ap.rst         |  492 +++++---
> >   MAINTAINERS                            |    2 +-
> >   drivers/s390/crypto/ap_bus.c           |   35 +-
> >   drivers/s390/crypto/vfio_ap_drv.c      |  124 +-
> >   drivers/s390/crypto/vfio_ap_ops.c      | 1436 ++++++++++++++++++------
> >   drivers/s390/crypto/vfio_ap_private.h  |   47 +-
> >   7 files changed, 1648 insertions(+), 593 deletions(-)
> >   create mode 100644 Documentation/s390/vfio-ap-locking.rst  
> 
> Unless somebody disagrees, I think we will carry these patches via the s390 tree.

No objections. I failed to verify that everything is in great shape (no
deadlocks, no races, and the logic is implemented like designed) because
I didn't have the bandwidth. I did spend some time looking at the
locking in v19 and I didn't find any problems. The locking design is
IMHO viable. I think Jason did a full review and r-b-ed all the patches
or is at least very close to it. I don't think it makes any sense to wait
for me. Sorry I couldn't do as much as I would have liked to.

Regards,
Halil

Powered by blists - more mailing lists