Hi Daniel,
On Thu, Feb 08, 2024 at 04:52:33PM +0000, Daniel P. Berrangé wrote:
> Date: Thu, 8 Feb 2024 16:52:33 +0000
> From: "Daniel P. Berrangé" <berrange@xxxxxxxxxx>
> Subject: Re: [PATCH v8 00/21] Introduce smp.modules for x86 in QEMU
>
> On Fri, Feb 02, 2024 at 12:10:58AM +0800, Zhao Liu wrote:
> > Hi Daniel,
> >
> > On Thu, Feb 01, 2024 at 09:21:48AM +0000, Daniel P. Berrangé wrote:
> > > Date: Thu, 1 Feb 2024 09:21:48 +0000
> > > From: "Daniel P. Berrangé" <berrange@xxxxxxxxxx>
> > > Subject: Re: [PATCH v8 00/21] Introduce smp.modules for x86 in QEMU
> > >
> > > On Thu, Feb 01, 2024 at 10:57:32AM +0800, Zhao Liu wrote:
> > > > Hi Daniel,
> > > >
> > > > On Wed, Jan 31, 2024 at 10:28:42AM +0000, Daniel P. Berrangé wrote:
> > > > > Date: Wed, 31 Jan 2024 10:28:42 +0000
> > > > > From: "Daniel P. Berrangé" <berrange@xxxxxxxxxx>
> > > > > Subject: Re: [PATCH v8 00/21] Introduce smp.modules for x86 in QEMU
> > > > >
> > > > > On Wed, Jan 31, 2024 at 06:13:29PM +0800, Zhao Liu wrote:
> > > > > > From: Zhao Liu <zhao1.liu@xxxxxxxxx>
> > > >
> > > > [snip]
> > > >
> > > > > > However, after digging deeper into the description and use cases of
> > > > > > cluster in the device tree [3], I realized that the essential
> > > > > > difference between clusters and modules is that cluster is an extremely
> > > > > > abstract concept:
> > > > > > * Cluster supports nesting though currently QEMU doesn't support
> > > > > > nested cluster topology. However, modules will not support nesting.
> > > > > > * Also due to nesting, there is great flexibility in sharing resources
> > > > > > on clusters, rather than narrowing cluster down to sharing L2 (and
> > > > > > L3 tags) as the lowest topology level that contains cores.
> > > > > > * Flexible nesting of cluster allows it to correspond to any level
> > > > > > between the x86 package and core.
> > > > > >
> > > > > > Based on the above considerations, and in order to eliminate the naming
> > > > > > confusion caused by the mapping between general cluster and x86 module
> > > > > > in v7, we now formally introduce smp.modules as the new topology level.
> > > > >
> > > > > What is the Linux kernel calling this topology level on x86 ?
> > > > > It will be pretty unfortunate if Linux and QEMU end up with
> > > > > different names for the same topology level.
> > > > >
> > > >
> > > > Now Intel's engineers in the Linux kernel are starting to use "module"
> > > > to refer to this layer of topology [4] to avoid confusion, where
> > > > previously the scheduler developers referred to the share L2 hierarchy
> > > > collectively as "cluster".
> > > >
> > > > Looking at it this way, it makes more sense for QEMU to use the
> > > > "module" for x86.
> > >
> > > I was thinking specificially about what Linux calls this topology when
> > > exposing it in sysfs and /proc/cpuinfo. AFAICT, it looks like it is
> > > called 'clusters' in this context, and so this is the terminology that
> > > applications and users are going to expect.
> >
> > The cluster related topology information under "/sys/devices/system/cpu/
> > cpu*/topology" indicates the L2 cache topology (CPUID[0x4]), not module
> > level CPU topology (CPUID[0x1f]).
> >
> > So far, kernel hasn't exposed module topology related sysfs. But we will
> > add new "module" related information in sysfs. The relevant patches are
> > ready internally, but not posted yet.
> >
> > In the future, we will use "module" in sysfs to indicate module level CPU
> > topology, and "cluster" will be only used to refer to the l2 cache domain
> > as it is now.
>
> So, if they're distinct concepts both relevant to x86 CPUs, then from
> the QEMU POV, should this patch series be changing the -smp arg to
> allowing configuration of both 'clusters' and 'modules' for x86 ?
Though the previous versions use "clusters" parameter, they, like the
current "modules" version, are just to add a CPU topology level to the
x86 CPU.
>
> An earlier version of this series just supported 'clusters', and this
> changed to 'modules', but your description of Linux reporting both
> suggests QEMU would need both.
>
Given the cluster support for x86, i.e. the L2 cache topology support,
we want to introduce a different cache topology configuration way than
CPU topology and avoid using the "cluster" as cache topology name (this
avoids the confusion of -smp "clusters" which is a CPU topology since
ARM also just treats "cluster" as a CPU topology level in QEMU other
than cache topology level).
BTW, for cache topology, may I ask for your advice? Currently, I can
think of 2 options:
1. Hacked the -smp as:
-smp cpus=4,sockets=2,cores=2,threads=1, \
l3-cache=socket,l2-cache=core,l1-i-cache=core,l1-d-cache=core
For this way, I just parsed the extended -smp and store the cache
topology in such structue:
typedef struct CacheTopology {
CPUTopoLevel l1i;
CPUTopoLevel l1d;
CPUTopoLevel l2;
CPUTopoLevel l3;
} CacheTopology;
This way is just used for smp cache topology. For the heterogeneous/hybrid
cache topology, I think it can be expanded based on the QOM CPU topology
[4] as:
-accel kvm -cpu host \
-device cpu-socket,id=sock0 \
-device cpu-die,id=die0,parent=sock0 \
-device cpu-module,id=module0,parent=die0 \
-device cpu-module,id=module1,parent=die0 \
-device cpu-core,id=core0,parent=module0,nr-threads=2 \
-device cpu-core,id=core1,parent=module1,nr-threads=1 \
-device cpu-core,id=core2,parent=module1,nr-threads=1 \
-device cache,id=cache0,parent=die0,level=3,type=unified \
-device cache,id=cache1,parent=core0,level=2,type=unified \
-device cache,id=cache2,parent=core0,level=1,type=data \
-device cache,id=cache3,parent=core0,level=1,type=inst \
-device cache,id=cache4,parent=module1,level=2,type=unified \
-device cache,id=cache5,parent=core1,level=1,type=data \
-device cache,id=cache6,parent=core1,level=1,type=inst \
-device cache,id=cache5,parent=core2,level=1,type=data \
-device cache,id=cache6,parent=core2,level=1,type=inst \
In the module0, the l2 (x86's cluster) is shared at core0 (core level).
And in the module1, the l2 is shared for core1 and core 2 (at module
level).
[4]: https://lore.kernel.org/qemu-devel/20231130144203.2307629-1-zhao1.liu@xxxxxxxxxxxxxxx/
2. But recently I realized maybe there's another option, which is just
to introduce a new option "-cache" like "-numa" to configure cache
topology.
In "-cache", we could accept the CPU list as the parameter:
-cache cache,cacheid=0,level=2,type=unified,cpus=0-1 \
-cache cache,cacheid=1,level=2,type=unified,cpus=2-3 \
or CPU topology ids as the parameters:
-cache cache,cache-id=0,level=2,type=unified \
-cache cache,cache-id=1,level=2,type=unified \
-cache cpu,cache-id=0,socket-id=0,die-id=0,module-id=0,core-id=0 \
-cache cpu,cache-id=1,socket-id=0,die-id=0,module-id=1 \
Hmmm, Daniel, which of the above two options do you prefer?
Thanks,
Zhao
