Re: [PATCH 2/3] swiotlb: prefix dma_to_phys and phys_to_dma functions
From: Sinan Kaya
Date: Tue Mar 29 2016 - 08:57:58 EST
On 3/29/2016 8:44 AM, Stefano Stabellini wrote:
> Could you please explain what is the problem that you are trying to
> solve? In other words, what is the issue with assuming that physical
> address = dma address (and the current dma_to_phys and phys_to_dma
> static inlines) if no arm64 platforms violate it? That's pretty much
> what is done on x86 too (without X86_DMA_REMAP).
> If you want to make sure that the assumption is not violated, you can
> introduce a boot time check or a BUG_ON somewhere.
> If there is an arm64 platform with phys_addr != dma_addr, we need proper
> support for it. In fact even if there is an IOMMU on that platform, when
> running Xen on it, the IOMMU would be used by the hypervisor and Linux
> would still end up without it, using the swiotlb.
The problem is that device drivers are trying to use dma_to_phys and phys_to_dma
APIs to do address translation even though these two API do not exist in DMA mapping
layer. (see patch #1 and I made the same mistake in my HIDMA code)
Especially, when a device is behind an IOMMU; the DMA addresses are not equal
to physical addresses. Usage of dma_to_phys causes a crash on the system.
I'm trying to prefix the dma_to_phys and phys_to_dma API with swiotlb so that
we know what it is intended for and usage of these API in drivers need to be
discouraged in the future during code reviews.
Since I renamed the API, Robin Murphy made a suggestion to convert this
#define swiotlb_to_virt(addr) phys_to_virt((phys_addr_t)(addr))
just to reduce code clutter since we know swiotlb_dma_to_phys returns phys already
for ARM64 architecture.
I think we can do this exercise later. I'll undo this change for the moment.
Let's focus on the API rename.
I don't want this general purpose dma_to_phys API returning phys=dma value. This is
Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project