[libvirt] Availability of libvirt-4.6.0 Release Candidate 2

[Laine Stump]

(Let's see now.... which people were the ones who believe it's a huge
insult to Cc them in list replies, and which are the ones who appreciate
the Cc and use it as a flag to raise the visibility of the message in
their mail client ..... ?????. Ah, I give up, just leaving in the Cc's
and let the complaints fly)(For the record, I like explicit Cc's when
you want to make sure I see something - that sends a copy to my phone,
which can be mildly annoying if it's a series of 40 patches, but also
very effective in getting my attention :-P)

On 08/01/2018 06:40 AM, Michal Privoznik wrote:
> On 07/31/2018 07:19 PM, Ján Tomko wrote:
>> And I'm still unsure about leaving in
>> commit 55ce65646348884656fd7bf3f109ebf8f7603494
>>    qemu: Use the correct vm def on cold attach
>> https://libvirt.org/git/?p=libvirt.git;a=commitdiff;h=55ce6564634
>> Which means attach-device --live --config will attach an interface
>> with a different MAC address in live and persistent definition. Laine?

(Let's see, you said my name once in the Cc, and once here. You really
need to say it *three* times to get my attention ala Beetlejuice or
Biggy Smalls. (and it helps to recite it into a mirror in a dark room,
apparently)).

>>
> Well, It's not only MAC address that can change. The device address
> might change too. This points to a broader problem. When we are parsing
> a device XML we fill in the blanks in postParse callbacks. However,
> those look only at either live or at inactive XML. Not at both at the
> same time. So how can we fill in the blanks that would be valid for both
> XMLs?

Yeah, this has been a problem for a long time (it's also problematic
for, e.g. the index of PCI controllers, which can cascade into
differences in the PCI addresses of devices that are connected to those
controllers; similar problem for other types of controllers of course).
I think I even tried fixing it once, but the solution was inadequate.

The problem comes when someone uses a mixture of --live only, --config
only, and --live+--config device attaches/detaches. At the base of the
problem is the fact that the live and config domaindef objects are
stored completely independent of each other, and the code that assigns
PCI addresses "seeds" the list of in-use address from a single domaindef
object. Perhaps we need to merge these into a single object where each
device is marked as "live", "config" or "both" - then any new device
added would be guaranteed to not conflict with any existing device in
live or config - the unified device list would be used to determine what
addresses are used, and the new device would just be added once to one
domaindef object (just with live and config flags set) so there would be
no need to copy it.

Or, well..., after about 30 seconds of thought I realize this would lead
to different behavior if someone e.g. deleted a currently-active device
only from the config, then wanted to add that same device back again -
the re-added device would end up with a different PCI address. So maybe
that isn't the right solution either.

At any rate, there is no perfect solution in sight for the current
release, so the question is whether the new (bad) behavior is better or
worse than the old (also bad) behavior. My understanding is that the old
behavior could lead to a config that had two devices at the same PCI
address, which is definitely undesirable. The new behavior could lead to
the PCI address of a newly-added device being different the next time
the guest is shutdown and restarted. I would tend to prefer the latter,
with the caveat that this new behavior provides a config that works
(from libvirt's domain parsing POV), but might create a strange error in
the guest that would be extremely difficult to troubleshoot (especially
6 months from now after we've all forgotten about this patch (and
forgotten about the idea that a more complete fix was needed).

So I'm undecided about my opinion. And when undecided I tend toward
inaction. Now *that's* helpful, isn't it?