Re: [Qemu-devel] [PATCH v4 07/10] s390x/sclp: properly guard pci-specific functions

[Halil Pasic <pasic@linux.vnet.ibm.com>]

On 08/22/2017 03:24 PM, Cornelia Huck wrote:
> On Tue, 22 Aug 2017 14:58:37 +0200
> Halil Pasic <pasic@linux.vnet.ibm.com> wrote:
> 
>> On 08/22/2017 11:39 AM, Cornelia Huck wrote:
>>> On Tue, 22 Aug 2017 11:20:51 +0200
>>> Halil Pasic <pasic@linux.vnet.ibm.com> wrote:
>>>   
>>>> On 08/22/2017 10:39 AM, Cornelia Huck wrote:  
>>>>>> I'm fine either way. If I imagine having a lots of adapter types, then I
>>>>>> would expect a switch or a jumptable on the type before handling control
>>>>>> to the pci specific function. In this case statically not supported types
>>>>>> would probably get caught by the default branch of the switch and for a
>>>>>> jumptable it could even handle the dynamic case (based on the facilities)
>>>>>> trivially. In short both approaches can make sense.    
>>>>> I'm also wondering at the naming (the command sounds very
>>>>> pci-specific). I'd just stick with this approach (modulo a possible
>>>>> change of the response code, for which I need to rely on you guys).
>>>>>     
>>>>
>>>>
>>>> Well, the QEMU name of the command is misleading misleading. In the AR
>>>> it's called 'Configure I/O Adapter'. The PCI comes into the picture via
>>>> byte 8 of the SCCB, the so called adapter type. Valid values for the
>>>> adapter type are: 00-01 reserved; 02 PCI function; 03-FF reserved. So
>>>> at this point we only have PCI.  
>>>
>>> OK, misleading naming combined with missing documentation leads to
>>> confusion...
>>>
>>> So:
>>>
>>> - s/PCI/IOA/ for SCLP_CMDW_{CONFIGURE,DECONFIGURE}_PCI  
>> nod
>>
>>> - have a switch/case over byte 8 with only one case (pci)  
>>
>> Let's say some kind of a check for bit 8 is a good idea.
> 
> Yes.
> 
>>
>>> - move the pci feature check into the pci code(? - not sure)  
>>
>> Don't know. Architecturally I don't see any direct connection
>> between the pci feature and this command.
> 
> I'd either have the check in the pci case for the adapter type, or in
> the called function. It's probably cleaner to do the check before
> calling the pci function.
> 

nod

>>
>> The availability of SCLP_CMDW_{,DE}CONFIGURE_IOA is indicated
>> by the result of the read scp info command read info in
>> ReadInfo.facilities. I think we should assume that the read scp
>> info command is always there.
> 
> Sure. But see the question in my other mail regarding the sclp
> facilities bit (does it cover pci or adapter reconfiguration?)

It (SCLP_HAS_PCI_RECONFIG) exactly  covers adapter reconfiguration.
That's what I tried to say with the paragraph above.

> 
>>
>> I would appreciate someone with AR access double checking.
> 
> I'd have hoped you had AR access :p

Yes, my statements are based solely on my reading of the AR (me
still lacks druid-knowledge). What I've asked for is a second
opinion (because AR-s are a twisty maze).

> 
>>
>>>
>>> There's still the question of when this sclp command first became
>>> available...
>>>   
>>
>> I would argue that it should not be important for AR
>> compliance. Currently it operates only on PCI so I doubt it
>> pre-dates PCI. But I don't think the current implementation
>> is buggy because it offers the sclp command regardless
>> of the zPCI facility.
> 
> I'm just wondering if there's another facility we're missing. The zpci
> facility might imply presence of adapter reconfiguration, but are there
> other facilities implying that as well, or even a dedicated facility?

Yes. The SCLP facility with the facility code 33 (aka SCLP_HAS_PCI_RECONFIG).
It is the dedicated facility.

I don't think zPCI architecturally implies the presence of this SCLP
command. And logically I would say it's either the other way around
adapter reconfiguration implies zPCI (because otherwise adapter
reconfiguration would be completely useless) or bidirectional. 

> 
>>
>> I don't know where should I look for the historical details
>> which go beyond the AR.
> 
> If there is no independent facility, it is probably best to just always
> provide the command, but fail for pci adapter type if the zpci facility
> is off.

IMHO we should SCLP_RC_INVALID_SCLP_COMMAND iff we don't provide
SCLP_HAS_PCI_RECONFIG (which has bad name again s/PCI/IOA).

I don't know of any facility except basic SCLP on which 
SCLP_HAS_PCI_RECONFIG depends on. 

For me both presenting and not presenting SCLP_HAS_PCI_RECONFIG
to the guest (via read SCP info SCLP command) in the absence of
the zPCI feature makes sense. The late because of the likely historical
reasons, the former because I think the AR allows it and it gives us more
flexibility.

>