4:27 a.m.
On Thu, Oct 05, 2017 at 10:44:29AM +0200, Michal Privoznik wrote:
On 10/05/2017 10:10 AM, Daniel P. Berrange wrote:
> On Wed, Oct 04, 2017 at 08:31:36AM +0200, Martin Kletzander wrote:
>> On Tue, Oct 03, 2017 at 03:10:48PM +0100, Daniel P. Berrange wrote:
>>> On Tue, Oct 03, 2017 at 04:03:20PM +0200, Martin Kletzander wrote:
>>>> On Tue, Oct 03, 2017 at 02:53:46PM +0100, Daniel P. Berrange wrote:
>>>>> On Tue, Oct 03, 2017 at 02:11:44PM +0200, Martin Kletzander wrote:
>>>>>> On Tue, Oct 03, 2017 at 12:58:59PM +0200, Michal Privoznik
wrote:
>>>>>>> https://bugzilla.redhat.com/show_bug.cgi?id=1434451
>>>>>>>
>>>>>>> It comes handy for management application to be able to have
a
>>>>>>> per-device label so that it can uniquely identify devices it
>>>>>>> cares about. The advantage of this approach is that we
don't have
>>>>>>> to generate aliases at define time (non trivial amount of
work
>>>>>>> and problems). The only thing we do is parse the user
supplied
>>>>>>> UUID and format it back. For instance:
>>>>>>>
>>>>>>> <disk type='block' device='disk'>
>>>>>>> <driver name='qemu' type='raw'/>
>>>>>>> <source dev='/dev/HostVG/QEMUGuest1'/>
>>>>>>> <target dev='hda' bus='ide'/>
>>>>>>>
<uuid>1efaf08b-9317-4b0f-b227-912e4bd9f483</uuid>
>>>>>>> <address type='drive' controller='0'
bus='0' target='0' unit='0'/>
>>>>>>> </disk>
>>>>>>>
>>>>>>> Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com>
>>>>>>> ---
>>>>>>>
>>>>>>> This is just a very basic implementation. If I get a green
light on this, I can
>>>>>>> implement the feature further, i.e. allow device lookup on
the UUID. For
>>>>>>> instance:
>>>>>>>
>>>>>>> virsh domiftune fedora $UUID $bandwidth
>>>>>
>>>>> <snip>
>>>>>
>>>>> I'm thinking that part of the problem we're having with
agreeing how to
>>>>> deal with this RFE is that we're over-analysing semantics, by
wondering
>>>>> whether its a unique name or UUID, its relation to alias, whether it
has
>>>>> bearing on APIs.
>>>>>
>>>>> How about we change tack, and do what we did when we needed
application
>>>>> specific information at the top level - just declare a general
purpose
>>>>> <metadata> element and say it is a completely opaque blob.
Libvirt will
>>>>> *never* do anything with it, other than to preserve it exactly as
is.
>>>>> No API will ever use the metadata in any way. Libvirt will never try
to
>>>>> guarantee uniqueness of metadata for each device. It can be JSON or
a
>>>>> gziped microsoft word document for all we care. Entirely upto the
app
>>>>> developer to decide what metadata is saved and guarantee uniqueness
if
>>>>> they so desired.
>>>>>
>>>>
>>>> That is kind of what I was aiming for. But in order for it to be cleaner
and
>>>> easier to use from user as well (and not only mgmt apps) I thought the
metadata
>>>> would just be one identifier. If you want to store more metadata for
the
>>>> device, then you can do all that in the domain metadata and just
reference the
>>>> particular device using the identifier if mgmt app wants to do that.
>>>
>>> Yes that is certainly possible. The caveats are that we still need a unique
>>> identifier for the device, and the metadata update is not atomic wrt
>>> to device hotplug.
>>>
>>
>> Yes, well, our (libvirt) unique identifier is not going anywhere, so
>> that's OK, we'll be using what we have been until now.
>>
>>> The plus side of the global metadata is that we have APIs to update it
>>> on the fly already, and its fully namespaced to allow multiple independant
>>> data sets to be stored.
>>>
>>
>> Yes, exactly.
>>
>>> I don't think lack of atomicity is a big deal as you could order it so
that
>>> you update metadata before doing the hotplug. Then worst case you have a
>>> device mentioned in metadata that doesn't exist, which is easy enough to
>>> detect.
>>>
>>
>> Right, if you want metadata for device, then you'll just update
>> metadata, hotplug device, and if it failed you update the metadata once
>> more.
>>
>> So are we on the same page? By that I mean agreeing on any sane user-supplied
>> identifier that we'll not guarantee uniqueness for, and neither will we use
it
>> for anything for now? (We can deal with the issues regarding using it when
>> someone wants to actually implement it).
>
> Per my reply to the earlier patch series, I'm now inclined to say that we
> should
>
> - Allow the mgmt app to set the aliases upfront
> - Auto-fill missing aliases at XML define time
>
> it has some downsides, but all the other solutions we've discussed have
> their own downsides too. So on balance I think its not worth it to add
> a second identifier for each device, when we already have alias.
They are not the same thing. Our alias is an ID we pass to QEMU and we
are dependent on it in multiple ways, plus we generate it with some of
our rules in mind. Allowing users to modify this would add so many
possible problems that I would not consider next two releases stable. I
would also not be confident enough to tell someone "just use the alias"
because I see all the places where it's not treated as just another
configurable parameter. And there's no problem with that. But there
will be if we allow users to set an identifier that's heavily used for
internal purposes.
On the other hand it has benefits too. The same alias works throughout
the stack. You can use your strings where you could use aliases, etc.
So what if, just to ease the inclusion of this feature, we force very
strict rules on user-supplied aliases and maybe give them a prefix.
Let's say if you want to supply the alias yourself, it has to be in form
of:
"ua-[a-zA-Z0-9]*"
otherwise libvirt will generate the alias itself and the only thing we
guarantee is that our aliases will never start with 'ua-'?
Question is if we are confident enough that:
a) apps will provide unique aliases (since we'll be putting user input
onto qemu cmd line)
b) apps will provide only allowed characters in the alias (not every
character can be in id=, can it?)
But I think we still have not answered this question: what if we need to
change pattern by which we generate aliases in the future? On one hand,
an alias is just a string so the pattern should not matter. On the other
hand, that's not quite true. For instance, "pci.0" has a very special
meaning. IOW, if we now worry about users cutting off the branch they
are sitting on, this is like giving them flamethrower in fireworks
production hall.
Michal