From: Christoph Hellwig Subject: Re: [PATCH 1/3] dma-mapping: remove the default map_resource implementation Date: Mon, 14 Jan 2019 18:08:55 +0100 Message-ID: <20190114170855.GA7485@lst.de> References: <20190111181731.11782-1-hch@lst.de> <20190111181731.11782-2-hch@lst.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Content-Disposition: inline In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org To: Robin Murphy Cc: Christoph Hellwig , Pawel Osciak , Marek Szyprowski , Kyungmin Park , Niklas =?iso-8859-1?Q?S=F6derlund?= , Russell King , linux-kernel@vger.kernel.org, iommu@lists.linux-foundation.org, Mauro Carvalho Chehab , linuxppc-dev@lists.ozlabs.org, linux-arm-kernel@lists.infradead.org, linux-media@vger.kernel.org List-Id: iommu@lists.linux-foundation.org On Mon, Jan 14, 2019 at 01:12:33PM +0000, Robin Murphy wrote: > Ignoring the offset was kind of intentional there, because at the time I > was primarily thinking about it in terms of the Keystone 2 platform where > the peripherals are all in the same place (0-2GB) in both the bus and CPU > physical address maps, and only the view of RAM differs between the two > (2-4GB vs. 32-34GB). However, on something like BCM283x, the peripherals > region is also offset from its bus address in the CPU view, but at a > *different* offset relative to that of RAM. I was more thinking of the PCIe P2P case, where we need to apply a consistent offset to translate between the CPU and the bus view. But this isn't really used for PCIe P2P, so I guess keeping the original sematics might be a better idea. That being said the videobuf code seems to rely on these offsets, so we might be between a rock and a hard place. > Fortunately, I'm not aware of any platform which has a DMA engine behind an > IOMMU (and thus *needs* to use dma_map_resource() to avoid said IOMMU > blocking the slave device register reads/writes) and also has any nonzero > offsets, and AFAIK the IOMMU-less platforms above aren't using > dma_map_resource() at all, so this change shouldn't actually break > anything, but I guess we have a bit of a problem making it truly generic > and robust :( Note that we don't actually use the code in this patch for ARM/ARM64 platforms with IOMMUs, as both the ARM and the ARM64 iommu code have their own implementations of ->map_resource that actually program the iommu (which at least for the PCIe P2P case would be wrong). > Is this perhaps another shove in the direction of overhauling > dma_pfn_offset into an arbitrary "DMA ranges" lookup table? It is long overdue anyway. >> addr = ops->map_resource(dev, phys_addr, size, dir, attrs); > > Might it be reasonable to do: > > if (!dma_is_direct(ops) && ops->map_resource) > addr = ops->map_resource(...); > else > addr = dma_direct_map_resource(...); > > and avoid having to explicitly wire up the dma_direct callback elsewhere? No, I absolutely _want_ the callback explicitly wired up. That is the only way to ensure we actually intentionally support it and don't just get a default that often won't work. Same issue for ->mmap and ->get_sgtable.