| Message ID | 159744117377.39197.9319853595178174798.stgit@naples-babu.amd.com |
|---|---|
| State | New |
| Headers | show |
| Series | Modify AMD topology to use socket/dies/core/thread model | expand |
On Fri, 14 Aug 2020 16:39:33 -0500 Babu Moger <babu.moger@amd.com> wrote: > Update the EPYC topology to use socket/dies/core/thread model. The EPYC > model does not use the smp dies to build the topology. Instead, it uses > numa nodes to build the topology. Internally both are similar concept > which divides the cores on L3 boundary. Combining both into one terminology > makes it simple to program. > > Add a new check to error out when smp dies are not provided when EPYC > model is numa configured. Next task is to remove node_id, nr_nodes and > nodes_per_pkg from EPYC topology which will be done in next patch. > > Signed-off-by: Babu Moger <babu.moger@amd.com> > --- > hw/i386/x86.c | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/hw/i386/x86.c b/hw/i386/x86.c > index 67bee1bcb8..e90c42d2fc 100644 > --- a/hw/i386/x86.c > +++ b/hw/i386/x86.c > @@ -138,6 +138,14 @@ void x86_cpus_init(X86MachineState *x86ms, int default_cpu_version) > > /* Check for apicid encoding */ > if (cpu_x86_use_epyc_apic_id_encoding(ms->cpu_type)) { > + if ((ms->numa_state->num_nodes > 0) && > + ms->numa_state->num_nodes != (ms->smp.sockets * x86ms->smp_dies)) { this case is gated by (ms->numa_state->num_nodes > 0) so it won't work in case -smp dies=>1 but there is no -numa options at all we need to error out and ask to provide numa nodes corresponding to (ms->numa_state->num_nodes == 0) && (ms->smp.sockets * x86ms->smp_dies) or better alternative would be to enable autonuma when EPYC cpu is enabled, that will insure that this patch will work even if user hasn't specified -numa option, since it will create a single numa node automatically. The later will take care of (-smp 1,dies=1) case, which is broken due to lack of explicit -numa we end up with CPU_UNSET_NUMA_NODE_ID in CPUID_Fn8000001E_ECX. > + error_setg(&error_fatal, "Numa configuration here requires smp " > + "'dies' parameter. Configure the cpu topology properly " > + "with max_cpus = sockets * dies * cores * threads. Dies" > + " is equivalent to number of numa nodes in a socket."); > + return; > + } > x86_set_epyc_topo_handlers(ms); > } > >
> -----Original Message----- > From: Igor Mammedov <imammedo@redhat.com> > Sent: Wednesday, August 19, 2020 6:26 AM > To: Moger, Babu <Babu.Moger@amd.com> > Cc: pbonzini@redhat.com; rth@twiddle.net; ehabkost@redhat.com; > mst@redhat.com; marcel.apfelbaum@gmail.com; qemu-devel@nongnu.org > Subject: Re: [PATCH v4 2/3] hw/i386: Update the EPYC topology to use > socket/dies/core/thread model > > On Fri, 14 Aug 2020 16:39:33 -0500 > Babu Moger <babu.moger@amd.com> wrote: > > > Update the EPYC topology to use socket/dies/core/thread model. The > > EPYC model does not use the smp dies to build the topology. Instead, > > it uses numa nodes to build the topology. Internally both are similar > > concept which divides the cores on L3 boundary. Combining both into > > one terminology makes it simple to program. > > > > Add a new check to error out when smp dies are not provided when EPYC > > model is numa configured. Next task is to remove node_id, nr_nodes and > > nodes_per_pkg from EPYC topology which will be done in next patch. > > > > Signed-off-by: Babu Moger <babu.moger@amd.com> > > --- > > hw/i386/x86.c | 8 ++++++++ > > 1 file changed, 8 insertions(+) > > > > diff --git a/hw/i386/x86.c b/hw/i386/x86.c index > > 67bee1bcb8..e90c42d2fc 100644 > > --- a/hw/i386/x86.c > > +++ b/hw/i386/x86.c > > @@ -138,6 +138,14 @@ void x86_cpus_init(X86MachineState *x86ms, int > > default_cpu_version) > > > > /* Check for apicid encoding */ > > if (cpu_x86_use_epyc_apic_id_encoding(ms->cpu_type)) { > > + if ((ms->numa_state->num_nodes > 0) && > > > + ms->numa_state->num_nodes != (ms->smp.sockets * > > + x86ms->smp_dies)) { > this case is gated by (ms->numa_state->num_nodes > 0) so it won't work in case > -smp dies=>1 but there is no -numa options at all Looking back again, this check does not appear to be a good idea. The numa information is coming from ACPI and die information is coming from CPU internal hardware config. It may not match. I don't think this check will be good option. Sorry. I did not think about that earlier. > > we need to error out and ask to provide numa nodes corresponding to > (ms->numa_state->num_nodes == 0) && (ms->smp.sockets * x86ms- > >smp_dies) > > or better alternative would be to enable autonuma when EPYC cpu is enabled, > that will insure that this patch will work even if user hasn't specified -numa > option, since it will create a single numa node automatically. > > The later will take care of (-smp 1,dies=1) case, which is broken due to lack of > explicit -numa we end up with CPU_UNSET_NUMA_NODE_ID in > CPUID_Fn8000001E_ECX. > > > + error_setg(&error_fatal, "Numa configuration here requires smp " > > + "'dies' parameter. Configure the cpu topology properly " > > + "with max_cpus = sockets * dies * cores * threads. Dies" > > + " is equivalent to number of numa nodes in a socket."); > > + return; > > + } > > x86_set_epyc_topo_handlers(ms); > > } > > > >
diff --git a/hw/i386/x86.c b/hw/i386/x86.c index 67bee1bcb8..e90c42d2fc 100644 --- a/hw/i386/x86.c +++ b/hw/i386/x86.c @@ -138,6 +138,14 @@ void x86_cpus_init(X86MachineState *x86ms, int default_cpu_version) /* Check for apicid encoding */ if (cpu_x86_use_epyc_apic_id_encoding(ms->cpu_type)) { + if ((ms->numa_state->num_nodes > 0) && + ms->numa_state->num_nodes != (ms->smp.sockets * x86ms->smp_dies)) { + error_setg(&error_fatal, "Numa configuration here requires smp " + "'dies' parameter. Configure the cpu topology properly " + "with max_cpus = sockets * dies * cores * threads. Dies" + " is equivalent to number of numa nodes in a socket."); + return; + } x86_set_epyc_topo_handlers(ms); }
Update the EPYC topology to use socket/dies/core/thread model. The EPYC model does not use the smp dies to build the topology. Instead, it uses numa nodes to build the topology. Internally both are similar concept which divides the cores on L3 boundary. Combining both into one terminology makes it simple to program. Add a new check to error out when smp dies are not provided when EPYC model is numa configured. Next task is to remove node_id, nr_nodes and nodes_per_pkg from EPYC topology which will be done in next patch. Signed-off-by: Babu Moger <babu.moger@amd.com> --- hw/i386/x86.c | 8 ++++++++ 1 file changed, 8 insertions(+)