Re: [Qemu-devel] [PATCH 2/9] s390x: fix invalid use of cc 1 for SSCH

[Cornelia Huck]

On Thu, 7 Sep 2017 13:01:34 +0200
Halil Pasic <address@hidden> wrote:

> On 09/07/2017 10:02 AM, Dong Jia Shi wrote:
> > * Cornelia Huck <address@hidden> [2017-09-06 13:25:38 +0200]:
> >   
> >> On Wed, 6 Sep 2017 16:27:20 +0800
> >> Dong Jia Shi <address@hidden> wrote:
> >>  
> >>> * Halil Pasic <address@hidden> [2017-09-05 19:20:43 +0200]:
> >>>  
> >>>>
> >>>>
> >>>> On 09/05/2017 05:46 PM, Cornelia Huck wrote:    
> >>>>> On Tue, 5 Sep 2017 17:24:19 +0200
> >>>>> Halil Pasic <address@hidden> wrote:
> >>>>>     
> >>>>>> My problem with a program check (indicated by SCSW word 2 bit 10) is
> >>>>>> that, in my reading of the architecture, the semantic behind it is: The
> >>>>>> channel subsystem (not the cu or device) has detected, that the 
> >>>>>> the channel program (previously submitted as an ORB) is erroneous. 
> >>>>>> Which
> >>>>>> programs are erroneous is specified by the architecture. What we have
> >>>>>> here does not qualify.
> >>>>>>
> >>>>>> My idea was to rather blame the virtual hardware (device) and put no 
> >>>>>> blame
> >>>>>> on the program nor he channel subsystem. This could be done using 
> >>>>>> device
> >>>>>> status (unit check with command reject, maybe unit exception) or 
> >>>>>> interface
> >>>>>> check. My train of thought was, the problem is not consistent across a
> >>>>>> device type, so it has to be device specific.    
> >>>>>
> >>>>> Unit exception might be a better way to express what is happening here.
> >>>>> At least, it moves us away from cc 1 and not towards cc 3 :)
> >>>>>     
> >>>>
> >>>> I will do a follow up patch pursuing device exception.
> >>>>     
> >>>>>>
> >>>>>> Of course blaming the device could mislead the person encountering the
> >>>>>> problem, and make him believe it's an non-virtual hardware problem.
> >>>>>>
> >>>>>> About the misleading, I think the best we can do is log out a message
> >>>>>> indicating what really happened.    
> >>>>>
> >>>>> Just document it in the code? If it doesn't happen with Linux as a
> >>>>> guest, it is highly unlikely to be seen in the wild.
> >>>>>     
> >>>>
> >>>>
> >>>> Well we have two problems here:
> >>>> 1) Unit exception can be already defined by the device type for the
> >>>> command (reference: 
> >>>> http://publibfp.dhe.ibm.com/cgi-bin/bookmgr/BOOKS/dz9ar110/2.6.10?DT=19920904110920).
> >>>> I think this one is what you mean. And I agree that's best handled
> >>>> with comment in code.    
> >>> Using unit check, with bit 3 byte 0 of the sense data set to 1, to
> >>> indicate an 'Equipment check', sounds a bit more proper than unit
> >>> exception.  
> >>
> >> I don't agree: Equipment check sounds a lot more dire (and seems to
> >> imply a malfunction). I like unit exception better.  
> > Got the point. Fair enough!
> >   
> 
> I do see some benefit in doing unit check over unit exception. Just
> kept quite to see the discussion unfold. As already said, unit exception
> seems to be something reserved for the device type to define in a more
> or less arbitrary but unambiguous way. I agreed to use this, because
> I trust Connie's assessment about not really being used by the
> devices in the wild (obviously nothing changed here).
> 
> If we consider the semantic of unit check with command reject, it's
> a surprisingly good match: basically device detected a programming
> error (which can not be detected by the channel-subsystem because it
> is device (type) specific). For reference see:
> http://publibfp.dhe.ibm.com/cgi-bin/bookmgr/BOOKS/dz9ar110/2.7.2.1?DT=19920904110920
> 
> IMHO that's almost exactly what we have here: the channel-program
> is good from the perspective of the channel subsystem, but the device
> can't deal with it. So we would not lie that the device is at fault
> (was Connie's concern initially) but we would not lie about having
> a generally invalid channel program (was my concern).
> 
> So how about an unit check with a command reject? (The only problem
> I see is is on the device vs device type plane -- but that ain't better
> for unit exception.)

I don't know, it feels a bit weird if I look at the cases where I saw
command reject in the wild before, even if seems to agree with the
architecture... but just a gut feeling.