Re: [PATCH rdma-next] RDMA/mlx5: Set mkeys for dmabuf at PAGE_SIZE

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Jun 13, 2024 at 09:01:42PM +0300, Leon Romanovsky wrote:
> From: Chiara Meiohas <cmeiohas@nvidia.com>
> 
> Set the mkey for dmabuf at PAGE_SIZE to support any SGL
> after a move operation.
> 
> ib_umem_find_best_pgsz returns 0 on error, so it is
> incorrect to check the returned page_size against PAGE_SIZE

This commit message is not clear enough for something that need to be
backported:

RDMA/mlx5: Support non-page size aligned DMABUF mkeys

The mkey page size for DMABUF is fixed at PAGE_SIZE because we have to
support a move operation that could change a large-sized page list
into a small page-list and the mkey must be able to represent it.

The test for this is not quite correct, instead of checking the output
of mlx5_umem_find_best_pgsz() the call to ib_umem_find_best_pgsz
should specify the exact HW/SW restriction - only PAGE_SIZE is
accepted.

Then the normal logic for dealing with leading/trailing sub page
alignment works correctly and sub page size DMBUF mappings can be
supported.

This is particularly painful on 64K kernels.

Jason