On Thu, Apr 04, 2019 at 07:01:43PM +0300, Gal Pressman wrote: > On 04-Apr-19 18:47, Jason Gunthorpe wrote: > > On Thu, Apr 04, 2019 at 03:25:26PM +0000, Saleem, Shiraz wrote: > >>> Subject: Re: [PATCH rdma-next v1 2/6] RDMA/umem: Add API to find best driver > >>> supported page size in an MR > >>> > >>> On 19-Feb-19 16:57, Shiraz Saleem wrote: > >>>> This helper iterates through the SG list to find the best page size to > >>>> use from a bitmap of HW supported page sizes. Drivers that support > >>>> multiple page sizes, but not mixed pages in an MR can use this API. > >>>> > >>>> Suggested-by: Jason Gunthorpe <jgg@xxxxxxxx> > >>>> Reviewed-by: Michael J. Ruhl <michael.j.ruhl@xxxxxxxxx> > >>>> Signed-off-by: Shiraz Saleem <shiraz.saleem@xxxxxxxxx> > >>> > >>> I've tested this patch comparing our existing efa_cont_pages() implementation vs > >>> ib_umem_find_single_pg_size() running different test suites: > >> > >> Thanks for testing! > >> > >>> > >>> Supported page sizes are anything between 4k to 2G. > >>> I'm using page shift to compare results of both functions (ilog2 the return value of > >>> ib_umem_find_single_pg_size). > >>> > >>> When huge pages are disabled, in many cases efa_cont_pages() returns page shift > >>> of 13 where ib_umem_find_single_pg_size() returns 12. Didn't see a test where > >>> ib_umem_find_single_pg_size() returns anything other than 12 (PAGE_SIZE). > >> > >> I wonder if the checks in place to guarantee offset into large page > >> align for user-space virt buf and phy buf is downgrading the page > >> bit. And that EFA might not need it. > > > > EFA doesn't have actual MRs with virtual addreses, so it probably > > doesn't care. > > Can you please explain this statement? The usual need for the virtual address alignment stems from having HW that can process RDMA READ and RDMA WRITE operations that contain a virtual address on the wire. Since EFA doesn't have rkeys, it only needs to sort things out for lkey which is perhaps simpler Otherwise the HW needs a 64 bit addr on the DMA path to fix the misalignment of VA to page size. Jason
