[PATCH v3 0/1] arm64/dma-mapping: reduce an unnecessary conversion for coherent DMA

[PATCH v3 0/1] arm64/dma-mapping: reduce an unnecessary conversion for coherent DMA

[Zhen Lei]

Changelog:
v2 -> v3:
1. Just add some description for this patch.

v1 -> v2:
1. Give up removing the conversion, because of thoughtless, instead moved it
into branch if (!is_device_dma_coherent(dev)). Thanks for Catalin's detailed
explanation, I directly take some relies as comment in the code.

Zhen Lei (1):
  arm64/dma-mapping: reduce an unnecessary conversion for coherent DMA

 arch/arm64/mm/dma-mapping.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

-- 
2.5.0

[PATCH v3 1/1] arm64/dma-mapping: reduce an unnecessary conversion for coherent DMA

[Zhen Lei]

1.For coherent DMA
  In swiotlb_alloc_coherent, it directly return vaddr on success, and
pass vaddr to free_pages on failure. So, we can directly transparent pass
vaddr from __dma_free to swiotlb_free_coherent.
  According to my testing, it can save 8 clock cycles.

2.For non-coherent DMA.
  Keep no change for the original processing flow.
  Because all DDR memory mapped as cacheable by default. But for
non-coherent devices, both CPUs and devices should use NC(non-cacheable)
attributes to access memory, to keep consistency. So we can not directly
use vaddr retured by __dma_alloc_coherent, but should further remap the
physical memory as NC and return it. So in __dma_free, we first use the
NC-vaddr to unmap, then get vaddr base upon dma_handle and use it to free
memory back to buddy.

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
---
 arch/arm64/mm/dma-mapping.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
index a6e757c..ceb2018 100644
--- a/arch/arm64/mm/dma-mapping.c
+++ b/arch/arm64/mm/dma-mapping.c
@@ -187,16 +187,22 @@ 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(dma_to_phys(dev, dma_handle));
-
 	size = PAGE_ALIGN(size);

 	if (!is_device_dma_coherent(dev)) {
 		if (__free_from_pool(vaddr, size))
 			return;
 		vunmap(vaddr);
+
+		/*
+		 * For non-coherent DMA, the vaddr is not part of the linear
+		 * mapping as it has been remapped by __dma_alloc() via
+		 * dma_common_contiguous_remap(), hence for swiotlb freeing we
+		 * need the actual linear map address.
+		 */
+		vaddr = phys_to_virt(dma_to_phys(dev, dma_handle));
 	}
-	__dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
+	__dma_free_coherent(dev, size, vaddr, dma_handle, attrs);
 }

 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
--
2.5.0