Re: [PATCH] hw/arm/acpi: Pack the SRAT processors structure by node_id ascending order

[Igor Mammedov <imammedo@redhat.com>]

On Wed, 8 Jan 2020 04:02:10 +0000
"Zengtao (B)" <prime.zeng@hisilicon.com> wrote:

> > -----Original Message-----
> > From: Igor Mammedov [mailto:imammedo@redhat.com]
> > Sent: Tuesday, January 07, 2020 11:50 PM
> > To: Zengtao (B)
> > Cc: Michael S. Tsirkin; qemu-devel@nongnu.org; qemu-trivial@nongnu.org;
> > Shannon Zhao; Peter Maydell; qemu-arm@nongnu.org
> > Subject: Re: [PATCH] hw/arm/acpi: Pack the SRAT processors structure by
> > node_id ascending order
> > 
> > On Tue, 7 Jan 2020 10:29:22 +0000
> > "Zengtao (B)" <prime.zeng@hisilicon.com> wrote:
> >   
> > > > -----Original Message-----
> > > > From: Michael S. Tsirkin [mailto:mst@redhat.com]
> > > > Sent: Tuesday, January 07, 2020 5:33 PM
> > > > To: Zengtao (B)
> > > > Cc: qemu-devel@nongnu.org; qemu-trivial@nongnu.org; Shannon  
> > Zhao;  
> > > > Peter Maydell; Igor Mammedov; qemu-arm@nongnu.org
> > > > Subject: Re: [PATCH] hw/arm/acpi: Pack the SRAT processors structure  
> > by  
> > > > node_id ascending order
> > > >
> > > > On Tue, Jan 07, 2020 at 05:18:49PM +0800, Zeng Tao wrote:  
> > > > > When booting the guest linux with the following numa configuration:
> > > > > -numa node,node_id=1,cpus=0-3
> > > > > -numa node,node_id=0,cpus=4-7
> > > > > We can get the following numa topology in the guest system:
> > > > > Architecture:          aarch64
> > > > > Byte Order:            Little Endian
> > > > > CPU(s):                8
> > > > > On-line CPU(s) list:   0-7
> > > > > Thread(s) per core:    1
> > > > > Core(s) per socket:    8
> > > > > Socket(s):             1
> > > > > NUMA node(s):          2
> > > > > L1d cache:             unknown size
> > > > > L1i cache:             unknown size
> > > > > L2 cache:              unknown size
> > > > > NUMA node0 CPU(s):     0-3
> > > > > NUMA node1 CPU(s):     4-7
> > > > > The Cpus 0-3 is assigned with NUMA node 1 in QEMU while it get  
> > NUMA  
> > > > node  
> > > > > 0 in the guest.
> > > > >
> > > > > In fact, In the linux kernel, numa_node_id is allocated per the ACPI
> > > > > SRAT processors structure order,so the cpu 0 will be the first one to
> > > > > allocate its NUMA node id, so it gets the NUMA node 0.
> > > > >
> > > > > To fix this issue, we pack the SRAT processors structure in numa node  
> > id  
> > > > > order but not the default cpu number order.
> > > > >
> > > > > Signed-off-by: Zeng Tao <prime.zeng@hisilicon.com>  
> > > >
> > > >
> > > > Does this matter? If yes fixing linux to take node id from proximity
> > > > field in ACPI seems cleaner ...
> > > >  
> > >
> > > In fact, I just want to align the node_id concept in QEMU and Linux.
> > > If we fix the kernel side, we need to align with all platforms.
> > > i think maybe not a good idea.  
> > if linux makes up node ID's on it's own, it would be hard for it to
> > map SRAT entries to other tables that use proximity id as well.
> > 
> > So it would need to maintain map of [proximity id] -> [node id] (and reverse)
> > somewhere to resolve mappings on other tables.
> > If it doesn't do this then it's broken and works just by accident,
> > if it does the fix probably should be in that code and not in QEMU.
> >   
> 
> Hmm, the problem is how to understand the concept of node id.
> 1. In dts, there is node id. Both the QEMU and Linux can use it
> directly, so no conflict.
> 2. In ACPI, there is only proximity domain, but no node id, there
>  should be a mapping between them, while kernel and QEMU maintain
>  their own rule, and unfortunately they conflict with each other 
>  sometimes.
> 
> There is no specification to indicate what we should do to maintain the
> mapping, so it's hard to align the QEMU and Linux.
>
> Any suggestion, or we just accept it as a rule since it don't affect much?

If node id generation is driven by SRAT content, it might be reasonable to
ask for SRAT parser in kernel to create node ids using proximity value
instead of the order ACPI_SRAT_PROCESSOR_GICC structures are enumerated.
That way node id would match ACPI spec.

But even with that I'd wouldn't expect cpu ids match as its basically
arbitrary numbers on both sided. One would need to use arch specific ids
to reliably match cpus on both sides (MPIDR in ARM case or APICID in x86).

> > >  
> > > > > ---
> > > > >  hw/arm/virt-acpi-build.c | 23 +++++++++++++++--------
> > > > >  1 file changed, 15 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/hw/arm/virt-acpi-build.c b/hw/arm/virt-acpi-build.c
> > > > > index bd5f771..497192b 100644
> > > > > --- a/hw/arm/virt-acpi-build.c
> > > > > +++ b/hw/arm/virt-acpi-build.c
> > > > > @@ -520,7 +520,8 @@ build_srat(GArray *table_data, BIOSLinker  
> > > > *linker, VirtMachineState *vms)  
> > > > >      AcpiSystemResourceAffinityTable *srat;
> > > > >      AcpiSratProcessorGiccAffinity *core;
> > > > >      AcpiSratMemoryAffinity *numamem;
> > > > > -    int i, srat_start;
> > > > > +    int i, j, srat_start;
> > > > > +    uint32_t node_id;
> > > > >      uint64_t mem_base;
> > > > >      MachineClass *mc = MACHINE_GET_CLASS(vms);
> > > > >      MachineState *ms = MACHINE(vms);
> > > > > @@ -530,13 +531,19 @@ build_srat(GArray *table_data, BIOSLinker  
> > > > *linker, VirtMachineState *vms)  
> > > > >      srat = acpi_data_push(table_data, sizeof(*srat));
> > > > >      srat->reserved1 = cpu_to_le32(1);
> > > > >
> > > > > -    for (i = 0; i < cpu_list->len; ++i) {
> > > > > -        core = acpi_data_push(table_data, sizeof(*core));
> > > > > -        core->type = ACPI_SRAT_PROCESSOR_GICC;
> > > > > -        core->length = sizeof(*core);
> > > > > -        core->proximity =  
> > > > cpu_to_le32(cpu_list->cpus[i].props.node_id);  
> > > > > -        core->acpi_processor_uid = cpu_to_le32(i);
> > > > > -        core->flags = cpu_to_le32(1);
> > > > > +    for (i = 0; i < ms->numa_state->num_nodes; ++i) {
> > > > > +        for (j = 0; j < cpu_list->len; ++j) {  
> > > >
> > > > Hmm O(n ^2) isn't great ...  
> > > Good suggestion, 3Q.  
> > > >  
> > > > > +            node_id =  
> > cpu_to_le32(cpu_list->cpus[j].props.node_id);  
> > > > > +            if (node_id != i) {
> > > > > +                continue;
> > > > > +            }
> > > > > +            core = acpi_data_push(table_data, sizeof(*core));
> > > > > +            core->type = ACPI_SRAT_PROCESSOR_GICC;
> > > > > +            core->length = sizeof(*core);
> > > > > +            core->proximity = node_id;
> > > > > +            core->acpi_processor_uid = cpu_to_le32(j);
> > > > > +            core->flags = cpu_to_le32(1);
> > > > > +        }
> > > > >      }  
> > > >
> > > > is the issue arm specific? wouldn't it affect x86 too?
> > > >  
> > > Good question, I think it will affect x86, but I need to confirm.
> > >  
> > > > >      mem_base = vms->memmap[VIRT_MEM].base;
> > > > > --
> > > > > 2.8.1  
> > >  
>