Re: [PATCH v8 00/12] s390x: CPU Topology

[Janis Schoetterl-Glausch <scgl@linux.ibm.com>]

On 7/15/22 15:47, Pierre Morel wrote:
> 
> 
> On 7/15/22 11:31, Janis Schoetterl-Glausch wrote:
>> On 7/14/22 22:05, Pierre Morel wrote:
>>>
>>>
>>> On 7/14/22 20:43, Janis Schoetterl-Glausch wrote:
>>>> On 6/20/22 16:03, Pierre Morel wrote:
>>>>> Hi,
>>>>>
>>>>> This new spin is essentially for coherence with the last Linux CPU
>>>>> Topology patch, function testing and coding style modifications.
>>>>>
>>>>> Forword
>>>>> =======
>>>>>
>>>>> The goal of this series is to implement CPU topology for S390, it
>>>>> improves the preceeding series with the implementation of books and
>>>>> drawers, of non uniform CPU topology and with documentation.
>>>>>
>>>>> To use these patches, you will need the Linux series version 10.
>>>>> You find it there:
>>>>> https://lkml.org/lkml/2022/6/20/590
>>>>>
>>>>> Currently this code is for KVM only, I have no idea if it is interesting
>>>>> to provide a TCG patch. If ever it will be done in another series.
>>>>>
>>>>> To have a better understanding of the S390x CPU Topology and its
>>>>> implementation in QEMU you can have a look at the documentation in the
>>>>> last patch or follow the introduction here under.
>>>>>
>>>>> A short introduction
>>>>> ====================
>>>>>
>>>>> CPU Topology is described in the S390 POP with essentially the description
>>>>> of two instructions:
>>>>>
>>>>> PTF Perform Topology function used to poll for topology change
>>>>>       and used to set the polarization but this part is not part of this item.
>>>>>
>>>>> STSI Store System Information and the SYSIB 15.1.x providing the Topology
>>>>>       configuration.
>>>>>
>>>>> S390 Topology is a 6 levels hierarchical topology with up to 5 level
>>>>>       of containers. The last topology level, specifying the CPU cores.
>>>>>
>>>>>       This patch series only uses the two lower levels sockets and cores.
>>>>>            To get the information on the topology, S390 provides the STSI
>>>>>       instruction, which stores a structures providing the list of the
>>>>>       containers used in the Machine topology: the SYSIB.
>>>>>       A selector within the STSI instruction allow to chose how many topology
>>>>>       levels will be provide in the SYSIB.
>>>>>
>>>>>       Using the Topology List Entries (TLE) provided inside the SYSIB we
>>>>>       the Linux kernel is able to compute the information about the cache
>>>>>       distance between two cores and can use this information to take
>>>>>       scheduling decisions.
>>>>
>>>> Do the socket, book, ... metaphors and looking at STSI from the existing
>>>> smp infrastructure even make sense?
>>>
>>> Sorry, I do not understand.
>>> I admit the cover-letter is old and I did not rewrite it really good since the first patch series.
>>>
>>> What we do is:
>>> Compute the STSI from the SMP + numa + device QEMU parameters .
>>>
>>>>
>>>> STSI 15.1.x reports the topology to the guest and for a virtual machine,
>>>> this topology can be very dynamic. So a CPU can move from from one topology
>>>> container to another, but the socket of a cpu changing while it's running seems
>>>> a bit strange. And this isn't supported by this patch series as far as I understand,
>>>> the only topology changes are on hotplug.
>>>
>>> A CPU changing from a socket to another socket is the only case the PTF instruction reports a change in the topology with the case a new CPU is plug in.
>>
>> Can a CPU actually change between sockets right now?
> 
> To be exact, what I understand is that a shared CPU can be scheduled to another real CPU exactly as a guest vCPU can be scheduled by the host to another host CPU.

Ah, ok, this is what I'm forgetting, and what made communication harder,
there are two ways by which the topology can change:
1. the host topology changes
2. the vCPU threads are scheduled on another host CPU

I've been only thinking about the 2.
I assumed some outside entity (libvirt?) pins vCPU threads, and so it would
be the responsibility of that entity to set the topology which then is 
reported to the guest. So if you pin vCPUs for the whole lifetime of the vm
then you could do that by specifying the topology up front with -devices.
If you want to support migration, then the outside entity would need a way
to tell qemu the updated topology.
 
> 
>> The socket-id is computed from the core-id, so it's fixed, is it not?
> 
> the virtual socket-id is computed from the virtual core-id

Meaning cpu.env.core_id, correct? (which is the same as cpu.cpu_index which is the same as
ms->possible_cpus->cpus[core_id].props.core_id)
And a cpu's core id doesn't change during the lifetime of the vm, right?
And so it's socket id doesn't either.

> 
>>
>>> It is not expected to appear often but it does appear.
>>> The code has been removed from the kernel in spin 10 for 2 reasons:
>>> 1) we decided to first support only dedicated and pinned CPU> 2) Christian fears it may happen too often due to Linux host scheduling and could be a performance problem
>>
>> This seems sensible, but now it seems too static.
>> For example after migration, you cannot tell the guest which CPUs are in the same socket, book, ...,
>> unless I'm misunderstanding something.
> 
> No, to do this we would need to ask the kernel about it.

You mean polling /sys/devices/system/cpu/cpu*/topology/*_id ?
That should work if it isn't done to frequently, right?
And if it's done by the entity doing the pinning it could judge if the host topology change
is relevant to the guest and if so tell qemu how to update it.
> 
>> And migration is rare, but something you'd want to be able to react to.
>> And I could imaging that the vCPUs are pinned most of the time, but the pinning changes occasionally.
> 
> I think on migration we should just make a kvm_set_mtcr on post_load like Nico suggested everything else seems complicated for a questionable benefit.

But what is the point? The result of STSI reported to the guest doesn't actually change, does it?
Since the same CPUs with the same calculated socket-ids, ..., exist.
You cannot migrate to a vm with a different virtual topology, since the CPUs get matched via the cpu_index
as far as I can tell, which is the same as the core_id, or am I misunderstanding something?
Migrating the MTCR bit is correct, if it is 1 than there was a cpu hotplug that the guest did not yet observe,
but setting it to 1 after migration would we wrong if the STSI result would be the same.
> 
>