[libvirt] [PATCH 8/8] Revert "conf: Validate disk lun using correct types"

[Cole Robinson]

On 05/10/2016 03:30 AM, Ján Tomko wrote:
> On Mon, May 09, 2016 at 11:57:20AM -0400, Cole Robinson wrote:
>> On 05/09/2016 04:59 AM, Peter Krempa wrote:
>>> On Fri, May 06, 2016 at 11:19:05 -0400, John Ferlan wrote:
>>>>> ---
>>>>>  src/conf/domain_conf.c   | 22 ----------------------
>>>>>  tests/qemuxml2argvtest.c |  3 +--
>>>>>  2 files changed, 1 insertion(+), 24 deletions(-)
>>>>>
>>>>
>>>> There was a bz associated with that commit - that'll need to be
>>>> addressed in some manner...
>>>
>>> Well, the initial assesment of that BZ was wrong. This should have been
>>> fixed in virt manager at that point.
>>>
>>
>> That bz: https://bugzilla.redhat.com/show_bug.cgi?id=1201143
>>
>> Is loaded with GSS and other 'priority' tags. Maybe the original assessment of
>> the bug is wrong. But just reverting the commit with little mention of that
>> until John dug it up is not helpful IMO
>>
> 
> Neither is discussing a private BZ on a public list:
> "You are not authorized to access bug #1201143. To see this bug, you
> must first log in to an account with the appropriate permissions."
> 

Apologies, I didn't notice it was private.

>>>> While I understand your point here, the configuration didn't work - that
>>>> is it couldn't be started anyway so there could not be a domain running
>>>> with that configuration and thus it wouldn't disappear on a subsequent
>>>> reload, hence why checking the config and rejecting "earlier" seemed
>>>
>>> It does not kill any running domain, that's right. Defined domains still
>>> vanish after that commit if they were defined before. That is still
>>> unwanted.
>>>
>>
>> Yes, this is problematic, but then again has this issue bitten us in the year
>> since this was committed? We should still fix it, but it's not time
>> critical... Maybe we come up with a better solution.
>>
>>>> proper even though we hadn't rejected such a config when the
>>>> "mode='host'" was first implemented.
>>>
>>> Only when introducing a feature you are allowed to do a check that
>>> rejects parsing XML, afterwards, no such thins should be added.
>>>
>>
>> Right, these rules make technical sense, but are extremely difficult to audit,
>> and have proven hard to enforce. People wandering into the code may follow the
>> conventions of the code around them, and have no idea that adding a new
>> validation check is 'bad' when the pre-existing similar checks are 'good'
>> strictly based one when they were added.
>>
>> I think we need a better framework for this. I'll probably send a larger mail
>> at some point, but basically I think we should:
>>
>> - rename VIR_DOMAIN_DEF_PARSE_SKIP_OSTYPE_CHECKS to something generic like
>> VIR_DOMAIN_DEF_PARSE_SKIP_VALIDATION. The original flag is used at libvirtd
>> startup time to avoid the 'disappearing domain' problem for when a qemu is
>> uninstalled for example, but we still get that validation check for normal
>> runtime XML define
>>
> 
> The problem with skipping validation is that we will end up with invalid
> domains being defined, breaking assumptions in other parts of libvirt.
> That way we would have to duplicate the checks on domain startup too
> (which would not be such a problem if they were all in the same
> function).
> 
> For example:
> commit 21b316f4d351859d9ccbf8a20199f7e8707fd51d
>     qemu: error out on missing machine type in configs
> which I added after someone tried to put the machine XML directly in
> /etc/libvirt.
> 

I think those types of errors are actually going to be limited. I can't think
of many cases where we set defaults in the generic parser that would have to
be skipped on libvirtd load. The arch+ostype+virt+machine stuff is pretty
special in that the default setting is heavily intertwined with the validation.

But yes if someone went with my approach they would need to do so carefully.
Part of that would be moving all the setting of defaults out of the XML parser
as well and into an explicit PostParse type function, so it's at least easier
to audit.

> There's no point in allowing the user to (try to) start such a domain,
> but currently we treat such unvalidated XML the same as XML from a fresh
> define.
> 
> 
> I think I recall an attempt to introudce an 'invalid' state, where you
> could not start the domain, but it was editable by libvirt APIs.
> Unfortunately I cannot find it in the archives. Does anyone remember
> what happened to it?
> 

Last patches I found are:

https://www.redhat.com/archives/libvir-list/2015-December/msg00027.html

But I didn't read the discussion to see why it stalled

Thanks,
Cole