On Wed, Nov 28, 2018 at 6:09 PM Rob Clark robdclark@gmail.com wrote:
On Wed, Nov 28, 2018 at 2:39 AM Christoph Hellwig hch@infradead.org wrote:
/** dma_sync_sg_*() flush the physical pages, so point* sg->dma_address to the physical ones for the right behavior.*/for_each_sg(msm_obj->sgt->sgl, s, msm_obj->sgt->nents, i)sg_dma_address(s) = sg_phys(s);I'm sorry, but this is completely bogus and not acceptable.
The only place that is allowed to initialize sg_dma_address is dma_map_sg. If the default dma ops don't work for your setup we have major a problem and need to fix the dma api / iommu integration instead of hacking around it.
I agree that the dma/iommu integration is very problematic for drm (in particular, gpu drivers that use the iommu as an gpu mmu).. Really we need a way that a driver can opt-out of this, and access the cpu cache APIs directly, skipping the dma API entirely. But as it is, we've had to hack around the dma API. I'm not really sure this hack is any worse than abusing dma_(un)map_sg() for doing cache operations.
I probably should have paid more attention and nak'd the dma/iommu integration before it landed. But given that now we are stuck in this situation, while I'm certainly interested if anyone has some ideas about how to let drivers opt out of the dma/iommu integration and bypass the dma API layer, I'm ok with replacing a hack with a less-bad hack.
May I take it as a positive nod to respin the next version?
Regards Vivek
BR, -R