Re: [PATCH v9 12/22] s390: vfio-ap: sysfs interfaces to configure control domains

[Halil Pasic]

On 08/22/2018 09:16 PM, Tony Krowiak wrote:
> On 08/22/2018 01:11 PM, Halil Pasic wrote:
> > On 08/22/2018 05:48 PM, Christian Borntraeger wrote:
> > > On 08/22/2018 05:34 PM, Pierre Morel wrote:
> > > > On 22/08/2018 17:11, Christian Borntraeger wrote:
> > > > > On 08/22/2018 01:03 PM, Pierre Morel wrote:
> > > > > > > That's interesting.
> > > > > > >
> > > > > > > > IMHO this quote is quite a half-full half-empty cup one:
> > > > > > > > * it mandates the set of usage domains is a subset of the set
> > > > > > > > of the control domains, but
> > > > > > > > * it speaks of independent controls, namely about the 'usage domain index'
> > > > > > > > and the 'control domain index list' and makes the enforcement of the rule
> > > > > > > > a job of the administrator (instead of codifying it in the controls).
> > > > > > > I'm wondering if a configuration with a usage domain that is not also a
> > > > > > > control domain is rejected outright? Anybody tried that? :)
> > > > > >
> > > > > > Yes, and no it is not.
> > > > > > We can use a queue (usage domain) to a AP card for SHA-512 or RSA without
> > > > > > having to define the queue as a control domain.
> > > > >
> > > > > Huh? My HMC allows to add a domain as
> > > > > - control only domain
> > > > > - control and usage domain.
> > > > >
> > > > > But I am not able to configure a usage-only domain for my LPAR. That seems to match
> > > > > the current code, no?
> > > >
> > > > Yes, it may not be configurable by the HMC but if we start a guest with no control domain it is not a problem to access the hardware through the usage domain.
> > > >
> > > > I tested this a long time ago, but tested again today to be sure on my LPAR.
> > > >
> > > > AFAIU adding a control only domain and a control and usage domain
> > > > allows say:
> > > > control and usage domain 1
> > > > control only domain 2
> > > >
> > > > Allow to send a message to domain 2 using queue 1
> > > >
> > > > Allow also to send a domain modifying message to domain 1 using queue 1
> > > >
> > > > control domain are domain which are controlled
> > >
> > > So you have changed the code to not automatically make a usage domain a
> > > control domain in the bitfield (and you could still use it as a usage
> > > domain). Correct?
> >
> > I tested basically the same yesterday, with the same results.
> >
> > > I think this is probably expected. the "usage implies control" seems to
> > > be a convention implemented by HMC (lpar) and z/VM but millicode offers
> > > the bits to have usage-only domains. As LPAR and z/VM will always enable
> > > any usage-domain to also be a control domain we should do the same.
> >
> > I'm fine either way, but slightly prefer higher level management software
> > and not the kernel accommodating this convention.
> >
> > Please consider a quote from Harald's mail in another sub-thread
> >
> >
> > """
> > ... about control domains
> >
> > Talked with the s390 firmware guys. The convention that the control domain
> > mask is a superset of the usage domain mask is only true for 1st level guests.
> >
> > It is absolutely valid to run a kvm guest with restricted control domain
> > mask bitmap in the CRYCB. It is valid to have an empty control domain mask
> > and the guest should be able to run crypto CPRBs on the usage domain(s) without
> > any problems. However, nobody has tried this.
> > """
> >
> > I'm yet to get an explanation why was this convention established in the first
> > place. And I can not figure it out myself. For me a setup where I know that
> > the domains used by some guest can not be modified by the same guest makes
> > perfect sense. If I try to think in analogies, I kind of compare modification
> > (that is control domain) with write access, and usage (that is usage domain)
> > with read access to, let's say a regular file. For me, all options (rw, r, and w)
> > do make sense, and if I had to pick the one that makes the least sense I would
> > pick write only. The convention is in these terms making read-only illegal. But
> > should 'usage only domains' ever get identified as something somebody wants to do
> > we can just add an attribute for that. So I'm fine either way.
>
> One of the things I suggested in a private conversation with Christian earlier
> today was to provide an additional rw sysfs attribute - a boolean - that indicates
> whether all usage domains should also be control domains. The default could be
> true. This would allow one to configure guests with usage-only domains as well
> as satisfy the convention.

I prefer keeping the attributes as they are and adding a new let's say
(un)assign_usage_domain if the need arises over this boolean attribute
that changes how (un)assign_domain works.

Halil