Re: [PATCH v4 2/2] kvm: Use huge pages for DAX-backed files

[Dan Williams]

On Thu, Dec 12, 2019 at 9:39 AM Liran Alon <liran.alon@xxxxxxxxxx> wrote:
>
>
>
> > On 12 Dec 2019, at 18:54, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> >
> > On Thu, Dec 12, 2019 at 4:34 AM Liran Alon <liran.alon@xxxxxxxxxx> wrote:
> >>
> >>
> >>
> >>> On 11 Dec 2019, at 23:32, Barret Rhoden <brho@xxxxxxxxxx> wrote:
> >>>
> >>> This change allows KVM to map DAX-backed files made of huge pages with
> >>> huge mappings in the EPT/TDP.
> >>>
> >>> DAX pages are not PageTransCompound.  The existing check is trying to
> >>> determine if the mapping for the pfn is a huge mapping or not.  For
> >>> non-DAX maps, e.g. hugetlbfs, that means checking PageTransCompound.
> >>> For DAX, we can check the page table itself.
> >>
> >> For hugetlbfs pages, tdp_page_fault() -> mapping_level() -> host_mapping_level() -> kvm_host_page_size() -> vma_kernel_pagesize()
> >> will return the page-size of the hugetlbfs without the need to parse the page-tables.
> >> See vma->vm_ops->pagesize() callback implementation at hugetlb_vm_ops->pagesize()==hugetlb_vm_op_pagesize().
> >>
> >> Only for pages that were originally mapped as small-pages and later merged to larger pages by THP, there is a need to check for PageTransCompound(). Again, instead of parsing page-tables.
> >>
> >> Therefore, it seems more logical to me that:
> >> (a) If DAX-backed files are mapped as large-pages to userspace, it should be reflected in vma->vm_ops->page_size() of that mapping. Causing kvm_host_page_size() to return the right size without the need to parse the page-tables.
> >
> > A given dax-mapped vma may have mixed page sizes so ->page_size()
> > can't be used reliably to enumerating the mapping size.
>
> Naive question: Why donât split the VMA in this case to multiple VMAs with different results for ->page_size()?

Filesystems traditionally have not populated ->pagesize() in their
vm_operations, there was no compelling reason to go add it and the
complexity seems prohibitive.

> What you are describing sounds like DAX is breaking this callback semantics in an unpredictable manner.

It's not unpredictable. vma_kernel_pagesize() returns PAGE_SIZE. Huge
pages in the page cache has a similar issue.

> >> (b) If DAX-backed files small-pages can be later merged to large-pages by THP, then the âstruct pageâ of these pages should be modified as usual to make PageTransCompound() return true for them. Iâm not highly familiar with this mechanism, but I would expect THP to be able to merge DAX-backed files small-pages to large-pages in case DAX provides âstruct pageâ for the DAX pages.
> >
> > DAX pages do not participate in THP and do not have the
> > PageTransCompound accounting. The only mechanism that records the
> > mapping size for dax is the page tables themselves.
>
> What is the rational behind this? Given that DAX pages can be described with âstruct pageâ (i.e. ZONE_DEVICE), what prevents THP from manipulating page-tables to merge multiple DAX PFNs to a larger page?

THP accounting is a function of the page allocator. ZONE_DEVICE pages
are excluded from the page allocator. ZONE_DEVICE is just enough
infrastructure to support pfn_to_page(), page_address(), and
get_user_pages(). Other page allocator services beyond that are not
present.