Re: [Qemu-devel] [RFC PATCH 1/3] hw/arm/virt: add hotplug memory support

[address@hidden]

Hi,

Thank you for the alignment information and discussion, Igor and Peter!

I'll try to find out proper alignment by looking at the codes.

Best Regards,
Kwangwoo Lee

> -----Original Message-----
> From: Igor Mammedov [mailto:address@hidden
> Sent: Monday, August 01, 2016 6:15 PM
> To: Peter Maydell
> Cc: 이광우(LEE KWANGWOO) MS SW; Xiao Guangrong; Michael S. Tsirkin; Paolo 
> Bonzini; Richard Henderson;
> Eduardo Habkost; Shannon Zhao; Shannon Zhao; QEMU Developers; qemu-arm; 
> 정우석(CHUNG WOO SUK) MS SW; 김
> 현철(KIM HYUNCHUL) MS SW
> Subject: Re: [RFC PATCH 1/3] hw/arm/virt: add hotplug memory support
> 
> On Mon, 1 Aug 2016 09:13:34 +0100
> Peter Maydell <address@hidden> wrote:
> 
> > On 1 August 2016 at 08:46, Igor Mammedov <address@hidden> wrote:
> > > On Mon, 1 Aug 2016 00:35:35 +0000
> > > "address@hidden" <address@hidden> wrote:
> >
> > >> > > +/* 1GB alignment for hotplug memory region */
> > >> > > +#define ARCH_VIRT_HOTPLUG_MEM_ALIGN (1ULL << 30)
> > >> >
> > >> > Where does the 1GB alignment come from? Why do we need 1GB
> > >> > alignment for the base but only 1KB alignment for the size?
> > >>
> > >> The alignment value of hotplug_memory.base referred from i386 pc.c and 
> > >> ppc spapr.c.
> > >> Could you suggest a proper range for this?
> >
> > > Base alignment comes from max supported hugepage size,
> >
> > Max hugepage size for any host? (if so, should be defined
> > in a common header somewhere)
> > Max hugepage size for ARM hosts? (if so, why is TCG
> > different from KVM?, and should still be in a common
> > header somewhere)
> It's the same for TCG but it probably doesn't matter much there,
> main usage is to provide better performance with KVM.
> 
> So I'd say it's host depended (for x86 it's 1Gb),
> probably other values for ARM and PPC
> 
> >
> > > while
> > > size alignment should depend on backend's page size
> >
> > Which page size do you have in mind here? TARGET_PAGE_SIZE
> > is often not the right answer, since it doesn't
> > correspond either to the actual page size being used
> > by the host kernel or to the actual page size used
> > by the guest kernel...
> alignment comes from here: memory_region_get_alignment()
> 
> exec:c
>    MAX(page_size, QEMU_VMALLOC_ALIGN)
> so it's either backend's page size or a min chunk QEMU
> allocates memory to make KVM/valgrind/whatnot happy.
> 
> >
> > thanks
> > -- PMM