Question about ARM KVM stage2 contiguous bit support for contiguous hugetlb

Question about ARM KVM stage2 contiguous bit support for contiguous hugetlb

[Zhou Wang]

Hi,

Current KVM stage2 supports contiguous hugetlb, e.g.4KB basic page size, 64KB hugetlb.
However, contiguous bit is not set in PTE, so there is no hint for hardware to create one
TLB for a contiguous huge page.

Seems to support this feature. We should do:

1. In user_mem_abort, do not force to do PTE map for CONT_PMD_SHIFT/CONT_PTE_SHIFT,
   just map all contiguous pages at one time.

2. Add contiguous bit in above process.

3. Modify the permission fault related code. As in the first point, we change vma_pagesize
   actually, vma_pagesize will be not equal with fault_granule, we will fail to enter into
   kvm_pgtable_stage2_relax_perms.

4. The split of contiguous hugetlb should be considered. We need remove all related contiguous
   bits, once one PTE has been changed.

Note sure if I miss any point to implement this feature. And do we have any plan to support this?

Best,
Zhou

Re: Question about ARM KVM stage2 contiguous bit support for contiguous hugetlb

[Marc Zyngier]

On Fri, 13 Mar 2026 07:09:40 +0000,
Zhou Wang <wangzhou1@hisilicon.com> wrote:
> 
> And do we have any plan to support this?

Not really.

Configurations that are looking at extracting the best possible
performance use large mappings at S2 (hugetlbfs). I'm not convinced
that there is a strong use-case for dealing with the contiguous hint
at S2 when backed by anonymous memory. If you're not under memory
pressure, THP already gives you the uplift you want. If you are under
memory pressure, you're unlikely to get the contiguity you want.

Additionally, dealing with the contiguous hint is a real pain.

If you are undertaking this work, please make sure to provide
reproducible benchmarks and results that show a definite improvement
on the majority of workloads.

Thanks,

	M.

-- 
Jazz isn't dead. It just smells funny.

Re: Question about ARM KVM stage2 contiguous bit support for contiguous hugetlb

[Zhou Wang]

On 2026/3/13 15:49, Marc Zyngier wrote:
> On Fri, 13 Mar 2026 07:09:40 +0000,
> Zhou Wang <wangzhou1@hisilicon.com> wrote:
>>
>> And do we have any plan to support this?
> 
> Not really.
> 
> Configurations that are looking at extracting the best possible
> performance use large mappings at S2 (hugetlbfs). I'm not convinced
> that there is a strong use-case for dealing with the contiguous hint
> at S2 when backed by anonymous memory. If you're not under memory

We want use 64KB basic page, in this case we have 2M cont PTE and 512M
block page. However, the latter may waste more memory.

> pressure, THP already gives you the uplift you want. If you are under
> memory pressure, you're unlikely to get the contiguity you want.
> 
> Additionally, dealing with the contiguous hint is a real pain.
> 
> If you are undertaking this work, please make sure to provide
> reproducible benchmarks and results that show a definite improvement
> on the majority of workloads.

We have a initial test in hardware emulator, for lmbench lat_mem_rd 1G
256 stride，the cost of virtualization will change from 17% -> 3%, we do
the test using a VM with 4 vCPU / 8GB memory in 64KB basic page / 2M
contiguous hugetlb system.

Will do more tests in the real chip.

Best,
Zhou

> 
> Thanks,
> 
> 	M.
>