Message-ID: <20171206111546.2db47aaa.cohuck@redhat.com>
Date:   Wed, 6 Dec 2017 11:15:46 +0100
From:   Cornelia Huck <cohuck@...hat.com>
To:     Pierre Morel <pmorel@...ux.vnet.ibm.com>
Cc:     Tony Krowiak <akrowiak@...ux.vnet.ibm.com>,
        Harald Freudenberger <freude@...ux.vnet.ibm.com>,
        Christian Borntraeger <borntraeger@...ibm.com>,
        Martin Schwidefsky <schwidefsky@...ibm.com>, freude@...ibm.com,
        mjrosato@...ux.vnet.ibm.com, pasic@...ux.vnet.ibm.com,
        Boris Fiuczynski <fiuczy@...ux.vnet.ibm.com>,
        linux-s390@...r.kernel.org, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, heiko.carstens@...ibm.com,
        kwankhede@...dia.com, bjsdjshi@...ux.vnet.ibm.com,
        pbonzini@...hat.com, alex.williamson@...hat.com,
        alifm@...ux.vnet.ibm.com, qemu-s390x@...gnu.org,
        jjherne@...ux.vnet.ibm.com, thuth@...hat.com
Subject: Re: [RFC 19/19] s390/facilities: enable AP facilities needed by
 guest

On Wed, 6 Dec 2017 10:15:51 +0100
Pierre Morel <pmorel@...ux.vnet.ibm.com> wrote:

> On 05/12/2017 16:01, Tony Krowiak wrote:
> > On 12/05/2017 09:04 AM, Cornelia Huck wrote:  
> >> On Tue, 5 Dec 2017 08:52:57 +0100
> >> Harald Freudenberger <freude@...ux.vnet.ibm.com> wrote:
> >>  
> >>> On 12/02/2017 02:30 AM, Tony Krowiak wrote:  
> >>>> I agree with your suggestion that defining a new CPU model feature 
> >>>> is probably
> >>>> the best way to resolve this issue. The question is, should we 
> >>>> define a single
> >>>> feature indicating whether AP instructions are installed and set 
> >>>> features bits
> >>>> for the guest based on whether or not they are set in the linux 
> >>>> host, or should
> >>>> we define additional CPU model features for turning features bits on 
> >>>> and off?
> >>>> I guess it boils down to what behavior is expected for the AP bus 
> >>>> running on
> >>>> the linux guest. Here is a rundown of the facilities bits associated 
> >>>> with AP
> >>>> and how they affect the behavior of the AP bus:
> >>>>
> >>>> * STFLE.12 indicates whether the AP query function is available. If 
> >>>> this bit
> >>>>    is not set, then the AP bus scan will only test domains 0-15. For 
> >>>> example,
> >>>>    if adapters 4, 5, and 6 and domains 12 and 71 (0x47) are 
> >>>> installed, then AP
> >>>>    queues 04.0047, 05.0047 and 06.0047 will not be made available.  
> >>> STFLE 12 is the indication for Query AP Configuration Information 
> >>> (QCI) available.  
> >>>> * STFLE.15 indicates whether the AP facilities test function is 
> >>>> available. If
> >>>>    this bit is not set, then the CEX4, CEX5 and CEX6 device drivers 
> >>>> discovered
> >>>>    by the AP bus scan will not get bound to any AP device drivers. 
> >>>> Since the
> >>>>    AP matrix model supports only CEX4 and greater, no devices will 
> >>>> be bound
> >>>>    to any driver for a guest.  
> >>> This T-Bit extension to the TAPQ subfunction is a must have. When kvm 
> >>> only
> >>> supports CEX4 and upper then this bit could also act as the indicator 
> >>> for
> >>> AP instructions available. Of course if you want to implement pure 
> >>> virtual
> >>> full simulated AP without any real AP hardware on the host this bit 
> >>> can't
> >>> be the indicator.  
> >> It would probably make sense to group these two together. Or is there
> >> any advantage in supporting only a part of it?  
> > After thinking about this a little more, I've come to the conclusion that
> > all of this might be moot for the following reasons:
> > 
> > * If STFLE.12 is not set for the linux host, then AP bus scan running on
> >    the host will not detect any domains with a domain number higher than 
> > 15,
> >    so no AP queues with a queue index higher than 15 will be available to
> >    bind to the vfio_ap_matrix driver. Consequently, no domain higher than
> >    15 can be assigned to any guest. In this case, the AP bus scan 
> > running on
> >    the guest will never detect a domain higher than 15, regardless of the
> >    setting of STFLE.12 for the guest.
> > 
> > * If STFLE.15 is not set for the linux host, then then there will be no
> >    CEX4, CEX5 or CEX6 queues available to bind to the vfio_ap_matrix
> >    driver, so no AP adapters or domains can be assigned to any KVM guest.
> > 
> > The bottom line is the STFLE bit settings for the linux host will control
> > what APs are available to the KVM guest. Since STFLE.15 controls whether
> > any CEX4,5 or 6 devices are even available, I think this bit can be
> > combined into the feature that indicates whether AP is available. As long
> > as AP instructions are available on the linux host, I'm not sure whether
> > STFLE.12 needs a feature at all.  
> 
> We are implementing VFIO with SIE interpretation.
> 
> 1) Providing more:
> The simple way is to provide to the guest only features existing on the 
> host.
> If we do provide features not existing on the host we need to be able to 
> emulate them.
> Even it is possible, it could be done in a future enhancement, but AFAIK 
> it is not the goal of the current development.

Yes. I think we currently want to provide a subset of what the SIE can
do. Any emulation would be icing on top.

> 
> 2) Providing less:
> On the other hand we can mask to the guest some of the features provided 
> by the host if we can intercept the scanning of the features.

Yes. I think that applies to bit 65 (interrupts) for now.

> 
> 
> What I understand from this is that we need all these features being 
> separately toggled to be able to be compatible with an older system even 
> if we have a 1:1 host:guest features match in a first version.

This seems to be the case for the mentioned bits if I followed the
discussion correctly.

> 
> If several features where introduced together in a new architecture and 
> are available on all systems issued from this architecture we can then 
> gather them in a set. (But I will wonder why we have several features then)

You're expecting that all architecture is making sense ;)

> 
> 
> >>  
> >>>> * STFLE.65 indicates whether AP interrupts are available. If this 
> >>>> bit is not
> >>>>    set, then the AP bus will use polling instead of using interrupt 
> >>>> handlers
> >>>>    to process AP events.  
> >> So, does this indicate "adapter interrupts for AP" only? If so, we
> >> should keep this separate and only enable it when we have the gisa etc.
> >> ready.  
> > Yes, this indicates AP interrupts only. The plan is to enable this when
> > GISA is available and we can implement interrupt processing.  
> 
> If we want to be able to work on system where STFLE.65 is not available, 
> even if GISA is available I think it would be interesting to have a 
> Matrix implementation with only polling.

Agreed.

So, it seems what we want is:

- A feature for STFLE 12. This seems to be a z13 or later facility, so
  it probably makes sense to indicate it in the guest for z13 or newer
  if the host supports it and for older machines if it is explicitly
  specified (and the host supports it).
- A feature for STFLE 15. Similar to the one above, but starting with a
  different generation (when was this introduced?)
- A feature for STFLE 65. Can be deferred to when GISA is implemented
  (and exploited by vfio-ap). Same as above (I think this has existed
  for a long time, probably for any of the machines we support?)

Thoughts?

Powered by blists - more mailing lists