Re: [PATCH for-6.2 v5 3/5] hw/arm/virt: Add cpu-map to device tree

["wangyanan (Y)" <wangyanan55@huawei.com>]

Hi,
On 2021/8/5 20:39, Yanan Wang wrote:
> From: Andrew Jones <drjones@redhat.com>
>
> Support device tree CPU topology descriptions.
>
> In accordance with the Devicetree Specification, the Linux Doc
> "arm/cpus.yaml" requires that cpus and cpu nodes in the DT are
> present. And we have already met the requirement by generating
> /cpus/cpu@* nodes for members within ms->smp.cpus. Accordingly,
> we should also create subnodes in cpu-map for the present cpus,
> each of which relates to an unique cpu node.
>
> The Linux Doc "cpu/cpu-topology.txt" states that the hierarchy
> of CPUs in a SMP system is defined through four entities and
> they are socket/cluster/core/thread. It is also required that
> a socket node's child nodes must be one or more cluster nodes.
> Given that currently we are only provided with information of
> socket/core/thread, we assume there is one cluster child node
> in each socket node when creating cpu-map.
>
> Signed-off-by: Andrew Jones <drjones@redhat.com>
> Co-developed-by: Yanan Wang <wangyanan55@huawei.com>
> Signed-off-by: Yanan Wang <wangyanan55@huawei.com>
> ---
>   hw/arm/virt.c | 59 ++++++++++++++++++++++++++++++++++++++++++---------
>   1 file changed, 49 insertions(+), 10 deletions(-)
>
> diff --git a/hw/arm/virt.c b/hw/arm/virt.c
> index 82f2eba6bd..d1e294be95 100644
> --- a/hw/arm/virt.c
> +++ b/hw/arm/virt.c
> @@ -350,20 +350,21 @@ static void fdt_add_cpu_nodes(const VirtMachineState *vms)
>       int cpu;
>       int addr_cells = 1;
>       const MachineState *ms = MACHINE(vms);
> +    const VirtMachineClass *vmc = VIRT_MACHINE_GET_CLASS(vms);
>       int smp_cpus = ms->smp.cpus;
>   
>       /*
> -     * From Documentation/devicetree/bindings/arm/cpus.txt
> -     *  On ARM v8 64-bit systems value should be set to 2,
> -     *  that corresponds to the MPIDR_EL1 register size.
> -     *  If MPIDR_EL1[63:32] value is equal to 0 on all CPUs
> -     *  in the system, #address-cells can be set to 1, since
> -     *  MPIDR_EL1[63:32] bits are not used for CPUs
> -     *  identification.
> +     * See Linux Documentation/devicetree/bindings/arm/cpus.yaml
> +     * On ARM v8 64-bit systems value should be set to 2,
> +     * that corresponds to the MPIDR_EL1 register size.
> +     * If MPIDR_EL1[63:32] value is equal to 0 on all CPUs
> +     * in the system, #address-cells can be set to 1, since
> +     * MPIDR_EL1[63:32] bits are not used for CPUs
> +     * identification.
>        *
> -     *  Here we actually don't know whether our system is 32- or 64-bit one.
> -     *  The simplest way to go is to examine affinity IDs of all our CPUs. If
> -     *  at least one of them has Aff3 populated, we set #address-cells to 2.
> +     * Here we actually don't know whether our system is 32- or 64-bit one.
> +     * The simplest way to go is to examine affinity IDs of all our CPUs. If
> +     * at least one of them has Aff3 populated, we set #address-cells to 2.
>        */
>       for (cpu = 0; cpu < smp_cpus; cpu++) {
>           ARMCPU *armcpu = ARM_CPU(qemu_get_cpu(cpu));
> @@ -406,8 +407,46 @@ static void fdt_add_cpu_nodes(const VirtMachineState *vms)
>                   ms->possible_cpus->cpus[cs->cpu_index].props.node_id);
>           }
>   
> +        if (!vmc->no_cpu_topology) {
> +            qemu_fdt_setprop_cell(ms->fdt, nodename, "phandle",
> +                                  qemu_fdt_alloc_phandle(ms->fdt));
> +        }
> +
>           g_free(nodename);
>       }
> +
> +    if (!vmc->no_cpu_topology) {
> +        /*
> +         * See Linux Documentation/devicetree/bindings/cpu/cpu-topology.txt
> +         * In a SMP system, the hierarchy of CPUs is defined through four
> +         * entities that are used to describe the layout of physical CPUs
> +         * in the system: socket/cluster/core/thread.
> +         */
> +        qemu_fdt_add_subnode(ms->fdt, "/cpus/cpu-map");
> +
> +        for (cpu = smp_cpus - 1; cpu >= 0; cpu--) {
> +            char *cpu_path = g_strdup_printf("/cpus/cpu@%d", cpu);
> +            char *map_path;
> +
> +            if (ms->smp.threads > 1) {
> +                map_path = g_strdup_printf(
> +                    "/cpus/cpu-map/socket%d/cluster0/core%d/thread%d",
> +                    cpu / (ms->smp.cores * ms->smp.threads),
> +                    (cpu / ms->smp.threads) % ms->smp.cores,
> +                    cpu % ms->smp.threads);
It seems that there is some discrepancy between the documentation
(Documentation/devicetree/bindings/cpu/cpu-topology.txt) and the
actual implementation of DT topology parser for ARM64
(function parse_dt_topology() in drivers/base/arch_topology.c).

The doc says the cpu-map node's child nodes can be:
     - one or more cluster nodes or
     - one or more socket nodes in a multi-socket system
which means a cpu-map can be defined as two formats such as:
1) cpu-map
                    socket0
                                 cluster0
                                              core0
                                              core1
                                 cluster1
                                              core0
                                              core1
                    socket1
                                 cluster0
                                              core0
                                              core1
                                 cluster1
                                              core0
                                              core1

2) cpu-map
                    cluster0
                                 cluster0
                                              core0
                                              core1
                                 cluster1
                                              core0
                                              core1
                    cluster1
                                 cluster0
                                              core0
                                              core1
                                 cluster1
                                              core0
                                              core1

But current parser only assumes that there are nested clusters within
cpu-map and is unaware of socket, the parser also ignore any information
about the nesting of clusters and present the scheduler with a flat list of
them. So based on current parser, we will get 4 packages (sockets) totally,
2 cores per package, 1 threads per core from 2), but will get nothing
useful from 1).

I think the ARM64 kernel DT parser should be optimized so that it's
also aware of sockets and can parse both formats of cpu-map. But
before this, I think we still have to build the cpu-map in format 2) if
we hope to describe topology successfully through DT. :)

Thanks,
Yanan
.
> +            } else {
> +                map_path = g_strdup_printf(
> +                    "/cpus/cpu-map/socket%d/cluster0/core%d",
> +                    cpu / ms->smp.cores,
> +                    cpu % ms->smp.cores);
> +            }
> +            qemu_fdt_add_path(ms->fdt, map_path);
> +            qemu_fdt_setprop_phandle(ms->fdt, map_path, "cpu", cpu_path);
> +
> +            g_free(map_path);
> +            g_free(cpu_path);
> +        }
> +    }
>   }
>   
>   static void fdt_add_its_gic_node(VirtMachineState *vms)