Re: [PATCH v6 21/21] s390: doc: detailed specifications for AP virtualization
From: Halil Pasic
Date: Tue Jul 03 2018 - 12:14:47 EST
On 07/03/2018 04:30 PM, Cornelia Huck wrote:
On Tue, 3 Jul 2018 15:58:37 +0200
Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:
On 07/03/2018 03:25 PM, Cornelia Huck wrote:
On Tue, 3 Jul 2018 14:20:11 +0200
Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:
On 07/03/2018 01:52 PM, Cornelia Huck wrote:
On Tue, 3 Jul 2018 11:22:10 +0200
Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:
[..]
Let me try to invoke the DASD analogy. If one for some reason wants to detach
a DASD the procedure to follow seems to be (see
https://www.ibm.com/support/knowledgecenter/en/linuxonibm/com.ibm.linux.z.lgdd/lgdd_t_dasd_online.html)
the following:
1) Unmount.
2) Offline possibly using safe_offline.
3) Detach.
Detaching a disk that is currently doing I/O asks for trouble, so the admin is encouraged
to make sure there is no pending I/O.
I don't think we can use dasd (block devices) as a good analogy for
every kind of device (for starters, consider network devices).
I did not use it for every kind of device. I used it for AP. I'm
under the impression you find the analogy inappropriate. If, could
you please explain why?
I don't think block devices (which are designed to be more or less
permanently accessed, e.g. by mounting a file system) have the same
semantics as ap devices (which exist as a backend for crypto requests).
Not everything that makes sense for a block device makes sense for
other devices as well, and I don't think it makes sense here.
I'm still confused. If it's about frequency of access (as hinted
by block devices accessed more or less permanently) I'm not sure
there is a substantial difference. I guess there are scenarios where
the AP domain is used very seldom (e.g. protected keys --> most of
the crypto ops done by CPACF but AP unwraps at the beginning), but
there are such scenarios for block too.
If it's about (persistent) state, I guess it again depends on the
scenario and on the type of the card. But I may be wrong.
So, let's turn this around: Why do you think that dasd (and not qeth or
whatever) is a good model for ap device unbinding? Because I really
fail to get it... maybe the ap driver maintainers can chime in.
Let's do it! But let me clarify one thing first I never stated that
dasd is the only good model.
What speaks for dasd as a model for unbinding:
* DASD is currently the only device we have vfio-mdev passthrough
for on s390x.
* DASD is comparatively simple and familiar. I'm not less confident
to talk about qeth or whatever else than to talk about DASD.
* DASD has persistent state. A NIC is much more stateless.
* DASD has offline and safe_offline. This kind of demonstrates that
the stock operation may trade 'safety' for stuff (e.g. guarantee to
terminate). Since the queue reset implemented by Tony has a limited
wait built in this seemed relevant.
* DASD can be seen as request-response with some local-ish stuff
as opposed to sending and receiving packets in a probably largish
network. The idea of outstanding operations is easy to gasp.
* From expectations of the upper layer entities a block device seems to
be a better fit than a network interface. Fault recovery is less of
a concern for an application that writes to a file, than for an
application that tires to talk to an other application over the net.
In my experience connections break more often that disks or I suppose
AP domains.
What is so wrong about asking the question: Is really unbind all
the admin has to do?
In case of AP you can interpret my 'in use' as the queue is not empty. In my understanding
unbind is supposed to be hard (I used the word radical). That's why I compared it to pulling
a cable. So that's why I ask is there stuff the admin is supposed to do before doing the
unbind.
Are you asking for a kind of 'quiescing' operation? I would hope that
the crypto drivers already can deal with that via flushing the queue,
not allowing new requests, or whatever. This is not the block device
case.
The current implementation of vfio-ap which is a crypto driver too certainly
can not deal 'with that'. Whether the rest of the drivers can, I don't
know. Maybe Tony can tell.
If the current implementation of vfio-ap cannot deal with it (by
cleaning up, blocking, etc.), it needs at the very least be documented
so that it can be implemented later. I do not know what the SIE will or
won't do to assist here (e.g., if you're removing it from some masks,
the device will already be inaccessible to the guest). But the part you
were referring to was talking about the existing host driver anyway,
wasn't it?
I was thinking about both directions. Re-classifying a device form
pass-through to normal should also be possible. But the document only
talks about one direction.
Presumably because it (rightfully) focuses on setting up vfio-ap?
I'm afraid we have a misunderstanding here. I did not propose to include
the other direction. Again I'm reasoning about the solution.
I'm not familiar with the existing host drivers. If we can say 'Hey,
unbind is perfectly safe at any time: no per-cautions need to be considered!'
I'm very happy with that. Although I would find it a bit surprising.
I just wanted to make sure this is not something we forget.
I'm aware of the fact that AP adapters are not block devices. But
as stated above I don't understand what is the big difference regarding
the unbind operation.
Anyway, this is an administrative issue. If you don't have a clear
concept which devices are for host usage and which for guest usage, you
already have problems.
I'm trying to understand the whole solution. I agree, this is an administrative
issue. But the document is trying to address such administrative issues.
I'd assume "know which devices are for the host and which devices are
for the guests" to be a given, no?
My other email scratches this topic. AFAIK we don't have a solution for
that yet. Nor we have a good understanding of how and to what extent
is statically given what is given. E.g. if one wants to re-partition my AP
resources (and at some point one will have to at least do the initial
re-partitioning) do I need a reboot for the changes to take effect? Or
is this 'known' variable during the uptime of an OS.
I think that is really out of scope for this file, which I'd expect to
explain how vfio-ap basically works and which incantations I need to
give crypto devices to a guest. It should NOT focus on administrative
tasks; this should either be delegated to the likes of libvirt or
documented in a "how to use crypto cards with kvm" kind of technical
writeup. If there's a limitation (e.g. you can't easily unbind again),
write a line here.
Again the misunderstanding. I'm not trying to understand the design and
not to put stuff in this document. I'm not aware of the existence of this
"how to use crypto cards with kvm" nor I've seen the likes of libvirt
patches that take care of the stuff. The stated purpose of this patch
is "provides documentation describing the AP architecture and
design concepts behind the virtualization of AP devices". This was the
best place I could find to ask my question. My intended question was
motivated by my understanding of unbind as a *not inherently safe*
operation, and by not knowing what happens if.
--
To unsubscribe from this list: send the line "unsubscribe linux-s390" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html