On Fri, Aug 2, 2019 at 2:37 PM Christoph Hellwig <h...@lst.de> wrote:
>
> Hi Rob,
>
> I just saw your above commit in Linus' tree, which is completely
> bogus and misunderstand the DMA API. Next time you have any issues
> please Cc the relevant maintainers and mailing lists. But even
> more importantly get_dma_ops is an completely internal API, and
> whatevet your helpers are trying to do is bad - if the dma wasn't
> mapped at the time you call the new sync_for_device helper, this
> is broken because you can't call dma_sync_sg_for_device on unmapped
> address. If it was mapped it is bogus as well as you can't double
> map it. Please describe what you're actually trying to fix and we're
> going to work on a proper solution, but this shot from the hip is just
> going to make things worse.
Sorry, this is just a temporary band-aid for v5.3 to get things
working again. Yes, I realize it is a complete hack.
The root problem is that I'm using the DMA API in the first place. I
don't actually use the DMA API to map buffers, for various reasons,
but instead manage the iommu_domain directly.
Because arm/arm64 cache ops are not exported to modules, so currently
I need to abuse the DMA API for cache operations (really just to clean
pages if I need to mmap them uncached/writecombine). Originally I was
doing that w/ dma_{map,unmap}_sg. But to avoid debug splats I
switched that to dma_sync_sg (see
0036bc73ccbe7e600a3468bf8e8879b122252274). But now it seems the
dma-direct ops are unhappy w/ dma_sync without a dma_map (AFAICT).
(On some generations of hw, the iommu is attached to the device node
that maps to the drm device, which is passed to dma_map/dma_sync. On
other generations the iommu is attached to a sub-device. Changing
this would break dtb compatibility.. so for now I need to handle both
iommu-ops and direct-ops cases.)
For v5.4, my plan was to revisit (at least for arm64) exposing cache
ops to drivers, so I can use drm_clflush_* instead, and stop abusing
DMA API.
BR,
-R
