Re: use of dma_direct_set_offset in (allwinner) drivers

[Maxime Ripard]

On Wed, Nov 04, 2020 at 10:15:49AM +0000, Robin Murphy wrote:
> On 2020-11-04 08:14, Maxime Ripard wrote:
> > Hi Christoph,
> > 
> > On Tue, Nov 03, 2020 at 10:55:38AM +0100, Christoph Hellwig wrote:
> > > Linux 5.10-rc1 switched from having a single dma offset in struct device
> > > to a set of DMA ranges, and introduced a new helper to set them,
> > > dma_direct_set_offset.
> > > 
> > > This in fact surfaced that a bunch of drivers that violate our layering
> > > and set the offset from drivers, which meant we had to reluctantly
> > > export the symbol to set up the DMA range.
> > > 
> > > The drivers are:
> > > 
> > > drivers/gpu/drm/sun4i/sun4i_backend.c
> > > 
> > >    This just use dma_direct_set_offset as a fallback.  Is there any good
> > >    reason to not just kill off the fallback?
> > > 
> > > drivers/media/platform/sunxi/sun4i-csi/sun4i_csi.c
> > > 
> > >    Same as above.
> > 
> > So, the history of this is:
> > 
> >    - We initially introduced the support for those two controllers
> >      assuming that there was a direct mapping between the physical and
> >      DMA addresses. It turns out it didn't and the DMA accesses were
> >      going through a secondary, dedicated, bus that didn't have the same
> >      mapping of the RAM than the CPU.
> > 
> >      4690803b09c6 ("drm/sun4i: backend: Offset layer buffer address by DRAM starting address")
> > 
> >    - This dedicated bus is undocumented and barely used in the vendor
> >      kernel so this was overlooked, and it's fairly hard to get infos on
> >      it for all the SoCs we support. We added the DT support for it
> >      though on some SoCs we had enough infos to do so:
> > 
> >      c43a4469402f ("dt-bindings: interconnect: Add a dma interconnect name")
> >      22f88e311399 ("ARM: dts: sun5i: Add the MBUS controller")
> > 
> >      This explains the check on the interconnect property
> > 
> >    - However, due to the stable DT rule, we still need to operate without
> >      regressions on older DTs that wouldn't have that property (and for
> >      SoCs we haven't figured out). Hence the fallback.
> 
> How about having something in the platform code that keys off the top-level
> SoC compatible and uses a bus notifier to create offsets for the relevant
> devices if an MBUS description is missing? At least that way the workaround
> could be confined to a single dedicated place and look somewhat similar to
> other special cases like sta2x11, rather than being duplicated all over the
> place.

I'll give it a try, thanks for the suggestion :)

Maxime

Attachment: signature.asc
Description: PGP signature