Re: [Qemu-devel] [PATCH] spapr: Enable use of huge pages

[Paolo Bonzini <pbonzini@redhat.com>]

Il 09/07/2014 07:57, Alexey Kardashevskiy ha scritto:
> 0b183fc87 "memory: move mem_path handling to
> memory_region_allocate_system_memory" disabled -mempath use for all
> machines that do not use memory_region_allocate_system_memory() to
> register RAM. Since SPAPR uses memory_region_init_ram(), the huge pages
> support was disabled for it.
>
> This replaces memory_region_init_ram()+vmstate_register_ram_global() with
> memory_region_allocate_system_memory() to get huge pages back.
>
> Cc: Paolo Bonzini <pbonzini@redhat.com>
> Cc: Hu Tao <hutao@cn.fujitsu.com>
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
>  hw/ppc/spapr.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> index a23c0f0..8fa9f7e 100644
> --- a/hw/ppc/spapr.c
> +++ b/hw/ppc/spapr.c
> @@ -1337,8 +1337,8 @@ static void ppc_spapr_init(MachineState *machine)
>          ram_addr_t nonrma_base = rma_alloc_size;
>          ram_addr_t nonrma_size = spapr->ram_limit - rma_alloc_size;
>
> -        memory_region_init_ram(ram, NULL, "ppc_spapr.ram", nonrma_size);
> -        vmstate_register_ram_global(ram);
> +        memory_region_allocate_system_memory(ram, NULL, "ppc_spapr.ram",
> +                                             nonrma_size);

The reason why I didn't do this in the simple way is that depending on 
the value of nonrma_base you may get smaller hugepages than you wanted.

For example, if the hugepage size is 1G but nonrma_base is 32M, you will 
not be able to get a page size larger than 32M.

Depending on the value of nonrma_base, it may be better to allocate the 
whole spapr->ram_limit to ppc_spapr.ram, and just ignore the first part 
of it.

I see in target-ppc/kvm.c that rma_alloc_size is capped to 256M, and  in 
practice it is 128M (arch/powerpc/kvm/book3s_hv_builtin.c.  Considering 
that Linux overcommits so the memory isn't lost in the non-hugepage 
case, I think it's better to just waste the 128M of address space.

Paolo

>          memory_region_add_subregion(sysmem, nonrma_base, ram);
>      }
>
>