"Kirill A. Shutemov" [off-list ref] writes:

On Thu, Mar 23, 2017 at 01:55:27PM -0700, Mike Kravetz wrote:

quoted

On 03/23/2017 05:58 AM, Punit Agrawal wrote:

quoted

On architectures that support hugepages composed of contiguous pte as
well as block entries at the same level in the page table,
huge_pte_offset() is not able to determine the right offset to return
when it encounters a swap entry (which is used to mark poisoned as
well as migrated pages in the page table).

huge_pte_offset() needs to know the size of the hugepage at the
requested address to determine the offset to return - the current
entry or the first entry of a set of contiguous hugepages. This came
up while enabling support for memory failure handling on arm64[0].

Patch 1 adds a hstate parameter to huge_pte_offset() to provide
additional information about the target address. It also updates the
signatures (and usage) of huge_pte_offset() for architectures that
override the generic implementation. This patch has been compile
tested on ia64 and x86.

I haven't looked at the performance implications of making huge_pte_offset
just a little slower.  But, I think you can get hstate from the parameters
passed today.

vma = find_vma(mm, addr);
h = hstate_vma(vma);

It's better to avoid find_vma() in fast(?) path if possible. So passing it
down is probably better.

Also most call sites of huge_pte_offset() already have the hstate (or
the vma) readily available. So adding overhead feels unnecessary.

I agree that merging the patch will need some co-ordination but lets at
the least give it a shot.