Re: [PATCH kvmtool] arm64: THPs for systems with pages bigger than 4K

[Vincent Donnefort <vdonnefort@google.com>]

On Tue, Oct 20, 2026 at 03:43:08PM +0100, Suzuki K Poulose wrote:
> On 19/10/2026 15:59, Vincent Donnefort wrote:
> > The THP support imposes an alignment of the memslot userspace address to
> > the size of a THP (PMD_SIZE). To do so, kvmtool overallocates the RAM
> > size and aligns it. However it assumes PMD_SIZE is 2MiB, which is only
> > true on 4KiB page systems.
> > 
> > Use a huge-page size, based on the system page size. This value can be
> > massive for a 64KiB page system, so skip the alignment if we couldn't
> > even install one THP.
> > 
> > Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
> > 
> > diff --git a/arm64/kvm.c b/arm64/kvm.c
> > index 36b3284e4a92..2c01dc32fa9a 100644
> > --- a/arm64/kvm.c
> > +++ b/arm64/kvm.c
> > @@ -18,6 +18,8 @@
> >   #include <asm/image.h>
> > +#define HUGEPAGE_SIZE (PAGE_SIZE * (PAGE_SIZE / 8UL))
> > +
> >   static struct arm64_image_header *kernel_header;
> >   struct kvm_ext kvm_req_ext[] = {
> > @@ -35,20 +37,23 @@ bool kvm__arch_cpu_supports_vm(void)
> >   void kvm__init_ram(struct kvm *kvm)
> >   {
> > -	u64 phys_start, phys_size;
> > +	u64 phys_start, phys_size, ram_alloc_align = SZ_64K;
> >   	void *host_mem;
> >   	int err;
> >   	/*
> > -	 * Allocate guest memory. We must align our buffer to 64K to
> > -	 * correlate with the maximum guest page size for virtio-mmio.
> > -	 * If using THP, then our minimal alignment becomes 2M.
> > -	 * 2M trumps 64K, so let's go with that.
> > +	 * Allocate guest memory. We must align our buffer to 64K to correlate
> > +	 * with the maximum guest page size for virtio-mmio. If using THP, then
> > +	 * our minimal alignment trumps 64K, so let's go with that.
> >   	 */
> >   	kvm->ram_size = kvm->cfg.ram_size;
> >   	kvm->arch.ram_alloc_size = kvm->ram_size;
> > -	if (!kvm->cfg.hugetlbfs_path)
> > -		kvm->arch.ram_alloc_size += SZ_2M;
> > +	if (!kvm->cfg.hugetlbfs_path) {
> > +		if (kvm->ram_size >= HUGEPAGE_SIZE)
> > +			ram_alloc_align = HUGEPAGE_SIZE;
> 
> There is logic in builtin-run.c to detect the host_ram_size(), could we
> use that instead to decide if we should apply this alignment ? Or retry
> the RAM allocation without the alignment on failure ?

As the mmap() sets MAP_NORESERVE, I believe it is only reserving address space,
I don't think a failure is expected because we increase the size of that
reservation or did I miss anything?

> 
> It doesn't look like a good idea to hard code this based just on the
> requested RAM size.

The idea is just if we know the guest doesn't have enough RAM to install at
least one THP, then there is no need over-mmap. But it isn't saving anything
really so perhaps I may as well drop it?

> 
> Cheers
> Suzuki
> 
> 

Thanks for having a look,

-- 
Vincent

> 
> > +
> > +		kvm->arch.ram_alloc_size += ram_alloc_align;
> > +	}
> >   	if (kvm->cfg.arch.guest_memfd)
> >   		kvm->arch.ram_alloc_start = mmap_guest_memfd(kvm,
> > @@ -63,7 +68,7 @@ void kvm__init_ram(struct kvm *kvm)
> >   		    kvm->arch.ram_alloc_size, errno);
> >   	kvm->ram_start = (void *)ALIGN((unsigned long)kvm->arch.ram_alloc_start,
> > -					SZ_2M);
> > +					ram_alloc_align);
> >   	if (!kvm->cfg.arch.guest_memfd) {
> >   		madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> > 
> > base-commit: f67bc0bdae9433a9cfd05e65ea2c1bb6102566d9
>