Re: [PATCH 03/15] KVM: selftests: Align HVA for HugeTLB-backed memslots

[Sean Christopherson]

On Thu, Feb 25, 2021, wangyanan (Y) wrote:
> Hi Sean,
> 
> On 2021/2/11 7:06, Sean Christopherson wrote:
> > Align the HVA for HugeTLB memslots, not just THP memslots.  Add an
> > assert so any future backing types are forced to assess whether or not
> > they need to be aligned.
> > 
> > Cc: Ben Gardon <bgardon@xxxxxxxxxx>
> > Cc: Yanan Wang <wangyanan55@xxxxxxxxxx>
> > Cc: Andrew Jones <drjones@xxxxxxxxxx>
> > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > Cc: Aaron Lewis <aaronlewis@xxxxxxxxxx>
> > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > ---
> >   tools/testing/selftests/kvm/lib/kvm_util.c | 5 ++++-
> >   1 file changed, 4 insertions(+), 1 deletion(-)
> > 
> > diff --git a/tools/testing/selftests/kvm/lib/kvm_util.c b/tools/testing/selftests/kvm/lib/kvm_util.c
> > index 584167c6dbc7..deaeb47b5a6d 100644
> > --- a/tools/testing/selftests/kvm/lib/kvm_util.c
> > +++ b/tools/testing/selftests/kvm/lib/kvm_util.c
> > @@ -731,8 +731,11 @@ void vm_userspace_mem_region_add(struct kvm_vm *vm,
> >   	alignment = 1;
> >   #endif
> > -	if (src_type == VM_MEM_SRC_ANONYMOUS_THP)
> > +	if (src_type == VM_MEM_SRC_ANONYMOUS_THP ||
> > +	    src_type == VM_MEM_SRC_ANONYMOUS_HUGETLB)
> Sorry for the late reply, I just returned from vacation.

At least you had an excuse :-)

> I am not sure HVA alignment is really necessary here for hugetlb pages.
> Different from hugetlb pages, the THP pages are dynamically allocated by
> later madvise(), so the value of HVA returned from mmap() is host page size
> aligned but not THP page size aligned, so we indeed have to perform
> alignment.  But hugetlb pages are pre-allocated on systems. The following
> test results also indicate that, with MAP_HUGETLB flag, the HVA returned from
> mmap() is already aligned to the corresponding hugetlb page size. So maybe
> HVAs of each hugetlb pages are aligned during allocation of them or in mmap()?

Yep, I verified the generic and arch version of hugetlb_get_unmapped_area() that
KVM supports all align the address.  PARISC64 is the only one that looks suspect,
but it doesn't support KVM.

> If so, I think we better not do this again here, because the later
> *region->mmap_size += alignment* will cause one more hugetlb page mapped but
> will not be used.

Agreed.  I think I'll still add the assert, along with a comment calling out
that HugetlB automatically handles alignment.

Nice catch, thanks!