Re: [PATCH v5 0/8] Remove EPYC mode apicid decode and use generic decode

[Igor Mammedov <imammedo@redhat.com>]

On Thu, 27 Aug 2020 15:07:52 -0400
Eduardo Habkost <ehabkost@redhat.com> wrote:

> On Thu, Aug 27, 2020 at 07:03:14PM +0200, Igor Mammedov wrote:
> > On Wed, 26 Aug 2020 16:03:40 +0100
> > Daniel P. Berrangé <berrange@redhat.com> wrote:
> > 
> > > On Wed, Aug 26, 2020 at 04:02:58PM +0200, Igor Mammedov wrote:
> > > > On Wed, 26 Aug 2020 14:36:38 +0100
> > > > Daniel P. Berrangé <berrange@redhat.com> wrote:
> > > > 
> > > > > On Wed, Aug 26, 2020 at 03:30:34PM +0200, Igor Mammedov wrote:
> > > > > > On Wed, 26 Aug 2020 13:50:59 +0100
> > > > > > Daniel P. Berrangé <berrange@redhat.com> wrote:
> > > > > >   
> > > > > > > On Wed, Aug 26, 2020 at 02:38:49PM +0200, Igor Mammedov wrote:  
> > > > > > > > On Fri, 21 Aug 2020 17:12:19 -0500
> > > > > > > > Babu Moger <babu.moger@amd.com> wrote:
> > > > > > > >     
> > > > > > > > > To support some of the complex topology, we introduced EPYC mode apicid decode.
> > > > > > > > > But, EPYC mode decode is running into problems. Also it can become quite a
> > > > > > > > > maintenance problem in the future. So, it was decided to remove that code and
> > > > > > > > > use the generic decode which works for majority of the topology. Most of the
> > > > > > > > > SPECed configuration would work just fine. With some non-SPECed user inputs,
> > > > > > > > > it will create some sub-optimal configuration.
> > > > > > > > > Here is the discussion thread.
> > > > > > > > > https://lore.kernel.org/qemu-devel/c0bcc1a6-1d84-a6e7-e468-d5b437c1b254@amd.com/
> > > > > > > > > 
> > > > > > > > > This series removes all the EPYC mode specific apicid changes and use the generic
> > > > > > > > > apicid decode.    
> > > > > > > > 
> > > > > > > > the main difference between EPYC and all other CPUs is that
> > > > > > > > it requires numa configuration (it's not optional)
> > > > > > > > so we need an extra patch on top of this series to enfoce that, i.e:
> > > > > > > > 
> > > > > > > >  if (epyc && !numa) 
> > > > > > > >     error("EPYC cpu requires numa to be configured")    
> > > > > > > 
> > > > > > > Please no. This will break 90% of current usage of the EPYC CPU in
> > > > > > > real world QEMU deployments. That is way too user hostile to introduce
> > > > > > > as a requirement.
> > > > > > > 
> > > > > > > Why do we need to force this ?  People have been successfuly using
> > > > > > > EPYC CPUs without NUMA in QEMU for years now.
> > > > > > > 
> > > > > > > It might not match behaviour of bare metal silicon, but that hasn't
> > > > > > > obviously caused the world to come crashing down.  
> > > > > > So far it produces warning in linux kernel (RHBZ1728166),
> > > > > > (resulting performance might be suboptimal), but I haven't seen
> > > > > > anyone reporting crashes yet.
> > > > > > 
> > > > > > 
> > > > > > What other options do we have?
> > > > > > Perhaps we can turn on strict check for new machine types only,
> > > > > > so old configs can keep broken topology (CPUID),
> > > > > > while new ones would require -numa and produce correct topology.  
> > > > > 
> > > > > No, tieing this to machine types is not viable either. That is still
> > > > > going to break essentially every single management application that
> > > > > exists today using QEMU.
> > > > for that we have deprecation process, so users could switch to new CLI
> > > > that would be required.
> > > 
> > > We could, but I don't find the cost/benefit tradeoff is compelling.
> > > 
> > > There are so many places where we diverge from what bare metal would
> > > do, that I don't see a good reason to introduce this breakage, even
> > > if we notify users via a deprecation message. 
> > I find (3) and (4) good enough reasons to use deprecation.
> > 
> > > If QEMU wants to require NUMA for EPYC, then QEMU could internally
> > > create a single NUMA node if none was specified for new machine
> > > types, such that there is no visible change or breakage to any
> > > mgmt apps.  
> > 
> > (1) for configs that started without -numa &&|| without -smp dies>1,
> >       QEMU can do just that (enable auto_enable_numa).
> 
> Why exactly do we need auto_enable_numa with dies=1?
> 
> If I understand correctly, Babu said earlier in this thread[1]
> that we don't need auto_enable_numa.
> 
> [1] https://lore.kernel.org/qemu-devel/11489e5f-2285-ddb4-9c35-c9f522d603a0@amd.com/

in case of 1 die, -numa is not must have as it's one numa node only.
Though having auto_enable_numa, will allow to reuse the CPU.node-id property
to compose CPUID_Fn8000001E_ECX. i.e only code one path vs numa|non-numa variant.

 
> > (2) As for configs that are out of spec, I do not care much (junk in - junk out)
> > (though not having to spend time on bug reports and debug issues, just to say
> > it's not supported in the end, makes deprecation sound like a reasonable
> > choice)
> > 
> > (3) However if config matches bare metal i.e. CPU has more than 1 die and within
> > dies limits (spec wise), QEMU has to produce valid CPUs.
> > In this case QEMU can't make up multiple numa nodes and mappings of RAM/CPUs
> > on user's behalf. That's where we have to error out and ask for explicit
> > numa configuration.
> > 
> > For such configs, current code (since 5.0), will produce in the best case
> > performance issues  due to mismatching data in APICID, CPUID and ACPI tables,
> > in the worst case issues might be related to invalid APIC ID if running on EPYC host
> > and HW takes in account subfields of APIC ID (according to Babu real CPU uses
> > die_id(aka node_id) internally).
> > I'd rather error out on nonsense configs earlier than debug such issues
> > and than error out anyways later (upsetting more users).
> > 
> 
> The requirements are not clear to me.  Is this just about making
> CPU die_id match the NUMA node ID, or are there additional
> constraints?
die_id is per socket numa node index, so it's not numa node id in
a sense we use it in qemu
(that's where all the confusion started that led to current code)

I understood that each die in EPYC chip is a numa node, which encodes
NUMA node ID (system wide) in CPUID_Fn8000001E_ECX, that's why I
wrote earlier that EPYC makes -numa non optional.

In case of only one die we can either use auto_enable_numa to ensure
that we have consistent code or special case it and just hardcode 
CPUID_Fn8000001E_ECX value which is hackish but will let us avoid
enabling numa (explicitly or implictly).

in case of multiple dies, CPUID_Fn8000001E_ECX (encodes number of nodes +
systemwide numa node id looking at CPUID of real EPYC machine)
shall match -numa mapping (otherwise it's a bug where CPUID and
ACPI mismatch).
Here we can go to ways:
  1) ask user to provide sane config with -numa (I'd prefer that)
     and use that info to fill in CPUID_Fn8000001E_ECX
  2) pretend that it's non numa machine, skip ACPI SRAT table
     but make up CPUID_Fn8000001E (i.e. another special case)
     (requires another code path and addition to -numa one)



> 
> 
> > (4)
> > If I were non hobby user, I'd hate if QEMU allowed me to start invalid config,
> > that I'd have to spend time on debugging issues (including performance ones),
> > instead of clearly telling me what's wrong and how config should be corrected.
> > I'd probably jump to another hypervisor that does the job right,
> > instead of digging into QEMU codebase and CPU specs to figure out how
> > to hack and configure it.
> > 
>