Re: [PATCH net-next v2 4/5] page_pool: disable sync for cpu for dmabuf memory provider

[Samiullah Khawaja]

On Thu, Nov 14, 2024 at 5:59 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Fri, Nov 08, 2024 at 11:01:21AM -0800, Mina Almasry wrote:
>
> > > If you do this you may want to block accepting dmabufs that have CPU
> > > pages inside them.

I believe we should be following the dmabuf API for this purpose, if
we really want to sync for CPU in. page_pool, and should not assume
the behaviour of the backing memory.

The dmabuf exporters are supposed to implement the begin_cpu_access
(optional) and the sync for cpu is done based on the exporter
implementation.

For example udmabuf implementation calls `dma_sync_sg_for_cpu` on the
underlying pages.

I think following dmabuf APIs can be used to ensure the backing memory
is synced:

Before cpu access to sync for cpu
`dma_buf_begin_cpu_access`

After cpu access to be synced for device
`dma_buf_end_cpu_access`

Sami
> > >
> >
> > How do I check if the dmabuf has CPU pages inside of it? The only way
> > I can think to do that is to sg_page a scatterlist entry, then
> > !is_zone_device_page() the page. Or something like that, but I thought
> > calling sg_page() on the dmabuf scatterlist was banned now.
>
> I don't know. Many dmabuf scenarios abuse scatter list and the CPU
> list is invalid, so you can't reference sg_page().
>
> I think you'd need to discuss with the dmabuf maintainers.
>
> Jason