Message-ID: <20171120181345.3fbda311.cohuck@redhat.com>
Date:   Mon, 20 Nov 2017 18:13:45 +0100
From:   Cornelia Huck <cohuck@...hat.com>
To:     Tony Krowiak <akrowiak@...ux.vnet.ibm.com>
Cc:     Pierre Morel <pmorel@...ux.vnet.ibm.com>,
        linux-s390@...r.kernel.org, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, freude@...ibm.com, schwidefsky@...ibm.com,
        heiko.carstens@...ibm.com, borntraeger@...ibm.com,
        kwankhede@...dia.com, bjsdjshi@...ux.vnet.ibm.com,
        pbonzini@...hat.com, alex.williamson@...hat.com,
        alifm@...ux.vnet.ibm.com, mjrosato@...ux.vnet.ibm.com,
        qemu-s390x@...gnu.org, jjherne@...ux.vnet.ibm.com,
        thuth@...hat.com, pasic@...ux.vnet.ibm.com
Subject: Re: [RFC 00/19] KVM: s390/crypto/vfio: guest dedicated crypto
 adapters

On Fri, 17 Nov 2017 15:28:16 -0500
Tony Krowiak <akrowiak@...ux.vnet.ibm.com> wrote:

> On 11/17/2017 05:07 AM, Cornelia Huck wrote:
> > On Fri, 17 Nov 2017 08:07:15 +0100
> > Pierre Morel <pmorel@...ux.vnet.ibm.com> wrote:
> >  
> >> On 17/11/2017 00:35, Tony Krowiak wrote:  
> >>> On 11/16/2017 03:25 PM, Pierre Morel wrote:  
> >>>> On 16/11/2017 18:03, Cornelia Huck wrote:  
> >>>>> On Thu, 16 Nov 2017 17:06:58 +0100
> >>>>> Pierre Morel <pmorel@...ux.vnet.ibm.com> wrote:

> >>>>>> So I totally agree with Conny on that we should stabilize the
> >>>>>> bus/device/driver modeling.
> >>>>>>
> >>>>>> I think it would be here a good place to start the discussion on things
> >>>>>> like we started to discuss, Harald and I, off-line:
> >>>>>> - why a matrix bus, in which case we can avoid it  
> >>>>> I thought it had been agreed that we should be able to ditch it?  
> >>>> I have not see any comment on the matrix bus.  
> >>> As stated in a previous email responding to Connie, I decided to scrap the
> >>> AP matrix bus. There will only ever be one matrix device that serves two
> >>> purposes: To hold the APQNs of the queue devices bound to the VFIO AP
> >>> matrix
> >>> device driver; to serve as a parent of the mediated devices created for
> >>> guests requiring access to the APQNs reserved for their use. So, instead
> >>> of an AP matrix bus creating the matrix device, it will be created by the
> >>> VFIO AP matrix driver in /sys/devices/ap_matrix/ during driver
> >>> initialization.  
> >> Sorry, I did not see the mail, this of course change a lot of things...  
> > One thing that would be useful for the next iteration is some ascii-art
> > representation that shows how the different parts (matrix, ap driver,
> > mdev, ...) tie together. That also would be useful to have in the
> > documentation.  
> I plan on including some drawings with the documentation and will include it
> in the cover letter as well.

Sounds good.

> >>>>>> - how to handle the repartition of queues on boot, reset and hotplug  
> >>> What do you mean by repartition of queues on boot?  
> >>>>> That's something I'd like to see a writeup for.  
> >>>> yes, and it may have an influence on the bus/device/driver/mdev design  
> >>> I don't understand the need to avoid implementation details. If you recall,
> >>> the original design was modeled on AP queue devices. It was only after
> >>> implementing that design that the shortcomings were revealed which is
> >>> why we decided to base the model on the AP matrix. Keep in mind, this is
> >>> an RFC, not a final patch set. I would expect some change from the
> >>> implementation herein. In fact, I've already made many changes based on
> >>> Connie's and Christian's review comments, none of which resulted in an
> >>> overhaul of the design.  
> > I expect that any of the above can be accommodated by the design. A
> > short writeup of what we may want to do for that would certainly help
> > to validate that, though.  
> I have spent some time thinking about hotplug implementation and I
> believe it can be accommodated within this design. I haven't looked
> at the implications for reset yet and I don't really know what is
> meant by "repartition of queues". I will include a write-up in the
> next submission.

FWIW, "repartition of queues" is also unclear to me.

> >>>>>> - interruptions  
> >>>>> My understanding is that interrupts are optional so they can be left
> >>>>> out in the first shot. With the gisa (that has not yet been posted), it
> >>>>> should not be too difficult, no?  
> >>>> you are right I forgot that it is optional  
> >>> If the facilities bit (STFLE.65) indicating interrupts are available is not
> >>> set for the guest, then the AP bus running on the guest will poll and
> >>> interrupts will not have to be handled. This patch set does not enable
> >>> interrupts, so it is not relevant at this time. We will not be able to
> >>> handle interrupts for the guest until the GISA for passthrough patches
> >>> are available. This will be addressed at that time.  
> > If you think it can be easily added later on, that would be fine for
> > me. (I cannot comment on gisa details until it has been posted,
> > obviously.)  
> Enabling AP interrupts is accomplished using the PQAP(AQIC) instruction
> which is a mandatory interception. The instruction will be forwarded to
> the VFIO AP device driver via an ioctl call on the mediated matrix
> device file descriptor. There will be some GISA set up needed and code
> to feed the interrupt back to user space, but I believe that will be
> provided by the forthcoming GISA passthrough patches. The bottom line is,
> I don't anticipate any major design change to handle interrupt processing.

Cool, that's what I wanted to hear.

> >>>>>> - virtualization of the AP  
> >>>>> Is this really needed? It would complicate everything a lot.  
> >>>> Concern has no sens without interception.  
> >>> Virtualization of AP is not on the table right now.  
> >> If we implement interception, we must speak about this, even to say how
> >> we do not implement virtualization.  
> > A note that we do not plan to virtualize it right now would be
> > sensible, yes.  
> Will do.
> >
> >  From what I remember, this would mean opening a huge can of worms for
> > something that might be only of limited use. I'd prefer a simplistic
> > but usable approach first. If virtualization should really become a
> > requirement in the future, it might be better served by a different
> > mechanism anyway.  
> I have done a little proof of concept code to get an idea if the AP matrix
> design will be extensible to handle virtualization. I modeled the
> proof of concept on the AP matrix model by creating a second mediated
> matrix device type for virtualization. Of course, virtual and passthrough
> matrix device types would have to be mutually exclusive; the admin would 
> have
> to choose one or the other. The sysfs model looked like this:
> 
> /sys/devices/ap_matrix
> ... [matrix]
> ...... [mdev_supported_types]
> ......... [vfio_ap_matrix-virtual]
> ............ create
> ............... [devices]
> .................. [$uuid]
> ..................... assign_adapter
> ..................... assign_domain
> 
> Using the a assign_adapter file, one can assign a virtual adapter
> ID to one or more real adapter IDs. For example, to assign virtual adapter
> 4 to real adapters 3, 22 and 254:
> 
> echo 4:3,22,254 > assign_adapter
> 
> Using the a assign_domain file, one can assign a virtual domain
> ID to one or more real domain IDs. For example, to assign virtual domain
> 0 to real domains 8 and 71:
> 
> echo 0:8,0x47 > assign_domain
> 
> All AP instructions would be intercepted for a virtual matrix. The 
> intercepted
> instructions would be forwarded to the VFIO AP matrix device driver by QEMU
> using an ioctl implemented by the VFIO AP matrix driver. If the mediated 
> matrix
> device is vfio_ap_matrix-passthrough type, things would work as they do now.
> If the type is vfio_ap_matrix-virtual, the the driver would:
> 
> 1. Calculate all of the real APQNs that can be used by:
>     * Retrieving the adapter IDs mapped to the APID specified in the APQN
>       contained in the AP instruction
>     * Retrieving the domain IDs mapped to the APQI specified in the APQN
>       contained in the AP instruction
>     * Combining all of the permutations of APID/APQI
> 2. Determine which APQN would be best to use.
> 3. Execute the instruction
> 4. Return the result to the caller
> 
> In other words, I think the current design is extensible; but even if not,
> I see no reason we can't design a completely different mechanism for
> virtualization.

So it's basically a one-time effort at (re)configuration, and the
virtualization facility will basically take care of the rest?

Powered by blists - more mailing lists