by Sinan Kaya

[permalink] [raw]

Subject: Re: [PATCH 2/3] swiotlb: prefix dma_to_phys and phys_to_dma functions

On 3/18/2016 8:12 AM, Robin Murphy wrote:
> Since we know for sure that swiotlb_to_phys is a no-op on arm64, it might be cleaner to simply not reference it at all. I suppose we could have some private local wrappers, e.g.:
>
> #define swiotlb_to_virt(addr) phys_to_virt((phys_addr_t)(addr))
>
> to keep the intent of the code clear (and just in case anyone ever builds a system mad enough to warrant switching out that definition, but I'd hope that never happens).
>
> Otherwise, looks good - thanks for doing this!

OK. I added this. Reviewed-by?

I'm not happy to submit such a big patch for all different ARCHs. I couldn't
find a cleaner solution. I'm willing to split this patch into multiple if there
is a better way.

diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
index ada00c3..8c0f66b 100644
--- a/arch/arm64/mm/dma-mapping.c
+++ b/arch/arm64/mm/dma-mapping.c
@@ -29,6 +29,14 @@

#include <asm/cacheflush.h>

+/*
+ * If you are building a system without IOMMU, then you are using SWIOTLB
+ * library. The ARM64 adaptation of this library does not support address
+ * translation and it assumes that physical address = dma address for such
+ * a use case. Please don't build a platform that violates this.
+ */
+#define swiotlb_to_virt(addr) phys_to_virt((phys_addr_t)(addr))
+
static pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot,
bool coherent)
{
@@ -188,7 +196,7 @@ static void __dma_free(struct device *dev, size_t size,
void *vaddr, dma_addr_t dma_handle,
struct dma_attrs *attrs)
{
- void *swiotlb_addr = phys_to_virt(swiotlb_dma_to_phys(dev, dma_handle));
+ void *swiotlb_addr = swiotlb_to_virt(dma_handle);

size = PAGE_ALIGN(size);

@@ -209,8 +217,7 @@ static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,

dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs);
if (!is_device_dma_coherent(dev))
- __dma_map_area(phys_to_virt(swiotlb_dma_to_phys(dev, dev_addr)),
- size, dir);
+ __dma_map_area(swiotlb_to_virt(dev_addr), size, dir);

return dev_addr;
}
@@ -283,8 +290,7 @@ static void __swiotlb_sync_single_for_device(struct device *dev,
{
swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
if (!is_device_dma_coherent(dev))
- __dma_map_area(phys_to_virt(swiotlb_dma_to_phys(dev, dev_addr)),
- size, dir);
+ __dma_map_area(swiotlb_to_virt(dev_addr), size, dir);
}

--
Sinan Kaya
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

2016-03-18 20:20:16

On Monday 28 March 2016 14:29:29 Konrad Rzeszutek Wilk wrote:
> > diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> > index ada00c3..8c0f66b 100644
> > --- a/arch/arm64/mm/dma-mapping.c
> > +++ b/arch/arm64/mm/dma-mapping.c
> > @@ -29,6 +29,14 @@
> >
> > #include <asm/cacheflush.h>
> >
> > +/*
> > + * If you are building a system without IOMMU, then you are using SWIOTLB
> > + * library. The ARM64 adaptation of this library does not support address
> > + * translation and it assumes that physical address = dma address for such
> > + * a use case. Please don't build a platform that violates this.
> > + */
>
> Why not just expand the ARM64 part to support address translation?

Because so far all hardware we have is relatively sane. We only
need to implement this if someone accidentally puts their DMA
space at the wrong address.

There is at least one platform that could in theory use this because
their RAM starts at an address that is outside of the reach of 32-bit
devices, and a static IOMMU mapping (created by firmware) could be
used to map the start of RAM into DMA address zero, to avoid having
to use an IOMMU all the time, but it was considered not worth the
effort to implement that.

Arnd