Re: [PATCH] lib/scatterlist: Provide a DMA page iterator

[Christian KÃnig]

Am 15.01.19 um 22:25 schrieb Jason Gunthorpe:
> On Tue, Jan 15, 2019 at 02:17:26PM +0000, Thomas Hellstrom wrote:
> > Hi, Christoph,
> >
> > On Mon, 2019-01-14 at 10:48 +0100, Christoph Hellwig wrote:
> > > On Thu, Jan 10, 2019 at 04:42:18PM -0700, Jason Gunthorpe wrote:
> > > > > Changes since the RFC:
> > > > > - Rework vmwgfx too [CH]
> > > > > - Use a distinct type for the DMA page iterator [CH]
> > > > > - Do not have a #ifdef [CH]
> > > > ChristophH: Will you ack?
> > > This looks generally fine.
> > >
> > > > Are you still OK with the vmwgfx reworking, or should we go back to
> > > > the original version that didn't have the type safety so this
> > > > driver
> > > > can be left broken?
> > > I think the map method in vmgfx that just does virt_to_phys is
> > > pretty broken.  Thomas, can you check if you see any performance
> > > difference with just doing the proper dma mapping, as that gets the
> > > driver out of interface abuse land?
> > The performance difference is not really the main problem here. The
> > problem is that even though we utilize the streaming DMA interface, we
> > use it only since we have to for DMA-Remapping and assume that the
> > memory is coherent. To be able to be as compliant as possible and ditch
> > the virt-to-phys mode, we *need* a DMA interface flag that tells us
> > when the dma_sync_for_xxx are no-ops. If they aren't we'll refuse to
> > load for now. I'm not sure, but I think also nouveau and radeon suffer
> > from the same issue.
> RDMA needs something similar as well, in this case drivers take a
> struct page * from get_user_pages() and need to have the DMA map fail
> if the platform can't DMA map in a way that does not require any
> additional DMA API calls to ensure coherence. (think Userspace RDMA
> MR's)
>
> Today we just do the normal DMA map and when it randomly doesn't work
> and corrupts data tell those people their platforms don't support RDMA
> - it would be nice to have a safer API base solution..

Oh, yes really good point. We have to support get_user_pages (or HMM) in 
a similar manner on GPUs as well.

Regards,
Christian.

> Jason
> _______________________________________________
> dri-devel mailing list
> dri-devel@xxxxxxxxxxxxxxxxxxxxx
> https://lists.freedesktop.org/mailman/listinfo/dri-devel