[PATCH v9 17/24] vfio/mlx5: Enable the DMA link API

[Leon Romanovsky]

On Wed, Apr 23, 2025 at 03:09:41PM -0300, Jason Gunthorpe wrote:
> On Wed, Apr 23, 2025 at 11:13:08AM +0300, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro at nvidia.com>
> > 
> > Remove intermediate scatter-gather table completely and
> > enable new DMA link API.
> > 
> > Tested-by: Jens Axboe <axboe at kernel.dk>
> > Signed-off-by: Leon Romanovsky <leonro at nvidia.com>
> > ---
> >  drivers/vfio/pci/mlx5/cmd.c  | 298 ++++++++++++++++-------------------
> >  drivers/vfio/pci/mlx5/cmd.h  |  21 ++-
> >  drivers/vfio/pci/mlx5/main.c |  31 ----
> >  3 files changed, 147 insertions(+), 203 deletions(-)
> 
> Reviewed-by: Jason Gunthorpe <jgg at nvidia.com>
> 
> > +static int register_dma_pages(struct mlx5_core_dev *mdev, u32 npages,
> > +			      struct page **page_list, u32 *mkey_in,
> > +			      struct dma_iova_state *state,
> > +			      enum dma_data_direction dir)
> > +{
> > +	dma_addr_t addr;
> > +	size_t mapped = 0;
> > +	__be64 *mtt;
> > +	int i, err;
> >  
> > -	return mlx5_core_create_mkey(mdev, mkey, mkey_in, inlen);
> > +	WARN_ON_ONCE(dir == DMA_NONE);
> > +
> > +	mtt = (__be64 *)MLX5_ADDR_OF(create_mkey_in, mkey_in, klm_pas_mtt);
> > +
> > +	if (dma_iova_try_alloc(mdev->device, state, 0, npages * PAGE_SIZE)) {
> > +		addr = state->addr;
> > +		for (i = 0; i < npages; i++) {
> > +			err = dma_iova_link(mdev->device, state,
> > +					    page_to_phys(page_list[i]), mapped,
> > +					    PAGE_SIZE, dir, 0);
> > +			if (err)
> > +				goto error;
> > +			*mtt++ = cpu_to_be64(addr);
> > +			addr += PAGE_SIZE;
> > +			mapped += PAGE_SIZE;
> > +		}
> 
> This is an area I'd like to see improvement on as a follow up.
> 
> Given we know we are allocating contiguous IOVA we should be able to
> request a certain alignment so we can know that it can be put into the
> mkey as single mtt. That would eliminate the double translation cost in
> the HW.
> 
> The RDMA mkey builder is able to do this from the scatterlist but the
> logic to do that was too complex to copy into vfio. This is close to
> being simple enough, just the alignment is the only problem.

I saw this improvement as well, but there is a need to generalize this 
"if (dma_iova_try_alloc) ... else ..." code first, as it will be used
by all vfio HW drivers.

So the plan is:
1. Merge the code as is.
2. Convert second vfio HW to the new API.
3. Propose something like dma_map_pages(..., struct page **page_list, ...)
   to map array of pages.
4. Optimize mlx5 vfio MTT creation.

Thanks

> 
> Jason
>