Re: [PATCH v2 1/4] xl: Add support for ignore_msrs option

[Jan Beulich <jbeulich@suse.com>]

On 22.02.2021 11:24, Roger Pau Monné wrote:
> On Fri, Feb 19, 2021 at 09:50:12AM -0500, Boris Ostrovsky wrote:
>>
>> On 2/18/21 10:57 AM, Jan Beulich wrote:
>>> On 18.02.2021 16:52, Roger Pau Monné wrote:
>>>> On Thu, Feb 18, 2021 at 12:54:13PM +0100, Jan Beulich wrote:
>>>>> On 18.02.2021 11:42, Roger Pau Monné wrote:
>>>>>> Not that you need to implement the full thing now, but maybe we could
>>>>>> have something like:
>>>>>>
>>>>>> "
>>>>>> =item B<ignore_msrs=[ "MSR_RANGE, "MSR_RANGE", ..]>
>>>>>>
>>>>>> Specify a list of MSR ranges that will be ignored by the hypervisor:
>>>>>> reads will return zeros and writes will be discarded without raising a
>>>>>> #GP.
>>>>>>
>>>>>> Each MSR_RANGE is given in hexadecimal format and may be a range, e.g.
>>>>>> c00102f0-c00102f1 (inclusive), or a single MSR, e.g. c00102f1.
>>>>>> "
>>>>>>
>>>>>> Then you can print the messages in the hypervisor using a guest log
>>>>>> level and modify it on demand in order to get more verbose output?
>>>>> "Modify on demand"? Irrespective of what you mean with this, ...
>>>>>
>>>>>> I don't think selecting whether the messages are printed or not from
>>>>>> xl is that helpful as the same could be achieved using guest_loglvl.
>>>>> ... controlling this via guest_loglvl would affect various other
>>>>> log messages' visibility.
>>>> Right, but do we really need this level of per-guest log control,
>>>> implemented in this way exclusively for MSRs?
>>
>>
>> In a multi-tenant environment we may need to figure out why a particular guest is failing to boot, without affecting behavior of other guests.
>>
>>
>> If we had per-guest log level in general then what you are suggesting would be the right thing to do IMO. Maybe that's what we should add?
> 
> Yes, that would seem better IMO, but I don't think it's fair to ask
> you to do that work.
> 
> Do you think it would be acceptable to untangle both, and try to get
> the MSR stuff without any logging changes?
> 
> I know we would be addressing only one part of what the series
> originally tried to achieve, but I would rather prefer to have a
> generic way to set a per-guest log level rather than something
> specific to MSR accesses.

TBH I'd see us go the other route: Follow Boris'es approach for
4.15, and switch the logging control to per-guest once that
ability is there, _and_ if we're really convinced we don't want
to have this extra level of control. The latter because I think
a domain could end up pretty chatty just because of MSR accesses,
and it might therefore be undesirable to also hide all other
potentially relevant output. Perhaps the per-domain log level
control needs to be finer grained than what "guest_loglvl="
currently permits, more like what "hvm_debug=" has.

Jan