Re: [PATCH] s390/vfio-ap: handle response code 01 on queue reset

[Halil Pasic]

On Tue, 05 Dec 2023 09:04:23 +0100
Harald Freudenberger <freude@xxxxxxxxxxxxx> wrote:

> On 2023-12-04 17:15, Halil Pasic wrote:
> > On Mon, 4 Dec 2023 16:16:31 +0100
> > Christian Borntraeger <borntraeger@xxxxxxxxxxxxx> wrote:
> >   
> >> Am 04.12.23 um 15:53 schrieb Tony Krowiak:  
> >> >
> >> >
> >> > On 11/29/23 12:12, Christian Borntraeger wrote:  
> >> >> Am 29.11.23 um 15:35 schrieb Tony Krowiak:  
> >> >>> In the current implementation, response code 01 (AP queue number not valid)
> >> >>> is handled as a default case along with other response codes returned from
> >> >>> a queue reset operation that are not handled specifically. Barring a bug,
> >> >>> response code 01 will occur only when a queue has been externally removed
> >> >>> from the host's AP configuration; nn this case, the queue must
> >> >>> be reset by the machine in order to avoid leaking crypto data if/when the
> >> >>> queue is returned to the host's configuration. The response code 01 case
> >> >>> will be handled specifically by logging a WARN message followed by cleaning
> >> >>> up the IRQ resources.
> >> >>>  
> >> >>
> >> >> To me it looks like this can be triggered by the LPAR admin, correct? So it
> >> >> is not desireable but possible.
> >> >> In that case I prefer to not use WARN, maybe use dev_warn or dev_err instead.
> >> >> WARN can be a disruptive event if panic_on_warn is set.  
> >> >
> >> > Yes, it can be triggered by the LPAR admin. I can't use dev_warn here because we don't have a reference to any device, but I can use pr_warn if that suffices.  
> >> 
> >> Ok, please use pr_warn then.  
> > 
> > Shouldn't we rather make this an 'info'. I mean we probably do not want
> > people complaining about this condition. Yes it should be a best 
> > practice
> > to coordinate such things with the guest, and ideally remove the 
> > resource
> > from the guest first. But AFAIU our stack is supposed to be able to
> > handle something like this. IMHO issuing a warning is excessive 
> > measure.
> > I know Reinhard and Tony probably disagree with the last sentence
> > though.  
> 
> Halil, Tony, the thing about about info versus warning versus error is 
> our
> own stuff. Keep in mind that these messages end up in the "debug 
> feature"
> as FFDC data. So it comes to the point which FFDC data do you/Tony want 
> to
> see there ? It should be enough to explain to a customer what happened
> without the need to "recreate with higher debug level" if something 
> serious
> happened. So my private decision table is:
> 1) is it something serious, something exceptional, something which may 
> not
>     come up again if tried to recreate ? Yes -> make it visible on the 
> first
>     occurrence as error msg.
> 2) is it something you want to read when a customer hits it and you tell 
> him
>     to extract and examine the debug feature data ? Yes -> make it a 
> warning
>     and make sure your debug feature by default records warnings.
> 3) still serious, but may flood the debug feature. Good enough and high
>     probability to reappear on a recreate ? Yes -> make it an info 
> message
>     and live with the risk that you may not be able to explain to a 
> customer
>     what happened without a recreate and higher debug level.
> 4) not 1-3, -> maybe a debug msg but still think about what happens when 
> a
>     customer enables "debug feature" with highest level. Does it squeeze 
> out
>     more important stuff ? Maybe make it dynamic debug with pr_debug() 
> (see
>     kernel docu admin-guide/dynamic-debug-howto.rst).

AFAIU the default log level of the S390 Debug Feature is 3 that is
error. So warnings do not help us there by default. And if we are 
already asking the reporter to crank up the loglevel of the debug
feature, we can as the reporter to crank it up to 5, assumed there
is not too much stuff that log level 5 in that area... How much
info stuff do we have for the 'ap' debug facility (I hope
that is the facility used by vfio_ap)? 

I think log levels are supposed to be primarily about severity, and
and I'm not sure that a queue becoming unavailable in G1 without
fist re-configuring the G2 so that it no more has access to the
given queue is not really a warning severity thing. IMHO if we
really do want people complaining about this should they ever see it,
yes it should be a warning. If not then probably not.

Regards,
Halil