Re: [Qemu-devel] [qemu-s390x] [PATCH v2 2/3] s390: cpu feature for diagnose 318 andlimit max VCPUs to 247

[Collin Walling <walling@linux.ibm.com>]

On 12/11/18 11:47 AM, Collin Walling wrote:
> On 12/7/18 7:08 AM, Cornelia Huck wrote:
>> On Thu,  6 Dec 2018 17:24:17 -0500
>> Collin Walling <walling@linux.ibm.com> wrote:
>>
>>> Diagnose 318 is a new z14.2 CPU feature. Since we are able to emulate
>>> it entirely via KVM, we can add guest support for earlier models. A
>>> new CPU feature for diagnose 318 (shortened to diag318) will be made
>>> available to guests starting with the zEC12-full CPU model.
>>>
>>> The z14.2 adds a new read SCP info byte (let's call it byte 134) to
>>> detect the availability of diag318. Because of this, we have room for
>>> one less VCPU and thus limit the max VPUs supported in a configuration
>>> to 247 (down from 248).
>>>
>>> Signed-off-by: Collin Walling <walling@linux.ibm.com>.
>>> ---
>>>  hw/s390x/sclp.c                 | 2 ++
>>>  include/hw/s390x/sclp.h         | 2 ++
>>>  target/s390x/cpu.h              | 2 +-
>>>  target/s390x/cpu_features.c     | 3 +++
>>>  target/s390x/cpu_features.h     | 1 +
>>>  target/s390x/cpu_features_def.h | 3 +++
>>>  target/s390x/gen-features.c     | 1 +
>>>  target/s390x/kvm.c              | 1 +
>>>  8 files changed, 14 insertions(+), 1 deletion(-)
>>>
>>
>>> diff --git a/target/s390x/cpu.h b/target/s390x/cpu.h
>>> index 8c2320e..594b4a4 100644
>>> --- a/target/s390x/cpu.h
>>> +++ b/target/s390x/cpu.h
>>> @@ -52,7 +52,7 @@
>>>  
>>>  #define MMU_USER_IDX 0
>>>  
>>> -#define S390_MAX_CPUS 248
>>> +#define S390_MAX_CPUS 247
>>
>> Isn't that already problematic if you try to migrate from an older QEMU
>> with all possible vcpus defined? IOW, don't you really need a way that
>> older machines can still run with one more vcpu?
>>
> 
> Good call. I'll run some tests on this and see what happens. I'll report
> here on those results.
> 

Migrating to a machine that supports less vCPUs will report

error: unsupported configuration: Maximum CPUs greater than specified machine type limit

I revisited the code to see if there's a way to dynamically set the max vcpu count based 
on the read scp info size, but it gets really tricky and code looks very complicated.
(Having a packed struct contain the CPU entries whose maximum is determined by hardware
limitations makes things difficult -- but who said s390 is easy? :) )

In reality, do we often have guests running with 248 or even 247 vcpus? If so, I imagine
the performance isn't too significant?

>>>  
>>>  typedef struct PSW {
>>>      uint64_t mask;
>>
> 
> 


-- 
Respectfully,
- Collin Walling