On Thu, Oct 31, 2019 at 1:38 AM Aneesh Kumar K.V
[off-list ref] wrote:

On 10/31/19 12:00 PM, Dan Williams wrote:

quoted

On Wed, Oct 30, 2019 at 10:35 PM Aneesh Kumar K.V
[off-list ref] wrote:
[..]

quoted

All that said, the x86 vmemmap_populate() falls back to use small
pages in some case to get around this constraint. Can't powerpc do the
same? It would seem to be less work than the above proposal.

ppc64 supports two translation mode (radix and hash). We can do the
above with radix translation. With hash we use just one page size( in
this specific case 16MB) for direct-map mapping.

Ok, if it's truly a hard constraint then yes, more infrastructure is
needed to expose that constraint to the namespace provisioning flow.