[PATCH v5] arm64: Add support for DMA_ATTR_FORCE_CONTIGUOUS to IOMMU

[geert+renesas@glider.be (Geert Uytterhoeven)]

Add support for allocating physically contiguous DMA buffers on arm64
systems with an IOMMU.  This can be useful when two or more devices
with different memory requirements are involved in buffer sharing.

Note that as this uses the CMA allocator, setting the
DMA_ATTR_FORCE_CONTIGUOUS attribute has a runtime-dependency on
CONFIG_DMA_CMA, just like on arm32.

For arm64 systems using swiotlb, no changes are needed to support the
allocation of physically contiguous DMA buffers:
  - swiotlb always uses physically contiguous buffers (up to
    IO_TLB_SEGSIZE = 128 pages),
  - arm64's __dma_alloc_coherent() already calls
    dma_alloc_from_contiguous() when CMA is available.

Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
---
v5:
  - Add Reviewed-by,
  - Pass GFP flags to dma_alloc_from_contiguous(), cfr. commit
    712c604dcdf818629 ("mm: wire up GFP flag passing in
    dma_alloc_from_contiguous"),

v4:
  - Replace dma_to_phys()/phys_to_page() by vmalloc_to_page(), to pass
    the correct page pointer to dma_release_from_contiguous().
    Note that the latter doesn't scream when passed a wrong pointer, but
    just returns false.  While this makes life easier for callers who
    may want to call another deallocator, it makes it harder catching
    bugs.

v3:
  - Add Acked-by,
  - Update comment to "one of _4_ things",
  - Call dma_alloc_from_contiguous() and iommu_dma_map_page() directly,
    as suggested by Robin Murphy,

v2:
  - New, handle dispatching in the arch (arm64) code, as requested by
    Robin Murphy.
---
 arch/arm64/mm/dma-mapping.c | 63 ++++++++++++++++++++++++++++++++++-----------
 1 file changed, 48 insertions(+), 15 deletions(-)

diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
index 81cdb2e844ed9fe8..f7b54019ef55378d 100644
--- a/arch/arm64/mm/dma-mapping.c
+++ b/arch/arm64/mm/dma-mapping.c
@@ -584,20 +584,7 @@ static void *__iommu_alloc_attrs(struct device *dev, size_t size,
 	 */
 	gfp |= __GFP_ZERO;
 
-	if (gfpflags_allow_blocking(gfp)) {
-		struct page **pages;
-		pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
-
-		pages = iommu_dma_alloc(dev, iosize, gfp, attrs, ioprot,
-					handle, flush_page);
-		if (!pages)
-			return NULL;
-
-		addr = dma_common_pages_remap(pages, size, VM_USERMAP, prot,
-					      __builtin_return_address(0));
-		if (!addr)
-			iommu_dma_free(dev, pages, iosize, handle);
-	} else {
+	if (!gfpflags_allow_blocking(gfp)) {
 		struct page *page;
 		/*
 		 * In atomic context we can't remap anything, so we'll only
@@ -621,6 +608,45 @@ static void *__iommu_alloc_attrs(struct device *dev, size_t size,
 				__free_from_pool(addr, size);
 			addr = NULL;
 		}
+	} else if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
+		pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
+		struct page *page;
+
+		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
+						 get_order(size), gfp);
+		if (!page)
+			return NULL;
+
+		*handle = iommu_dma_map_page(dev, page, 0, iosize, ioprot);
+		if (iommu_dma_mapping_error(dev, *handle)) {
+			dma_release_from_contiguous(dev, page,
+						    size >> PAGE_SHIFT);
+			return NULL;
+		}
+		if (!coherent)
+			__dma_flush_area(page_to_virt(page), iosize);
+
+		addr = dma_common_contiguous_remap(page, size, VM_USERMAP,
+						   prot,
+						   __builtin_return_address(0));
+		if (!addr) {
+			iommu_dma_unmap_page(dev, *handle, iosize, 0, attrs);
+			dma_release_from_contiguous(dev, page,
+						    size >> PAGE_SHIFT);
+		}
+	} else {
+		pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
+		struct page **pages;
+
+		pages = iommu_dma_alloc(dev, iosize, gfp, attrs, ioprot,
+					handle, flush_page);
+		if (!pages)
+			return NULL;
+
+		addr = dma_common_pages_remap(pages, size, VM_USERMAP, prot,
+					      __builtin_return_address(0));
+		if (!addr)
+			iommu_dma_free(dev, pages, iosize, handle);
 	}
 	return addr;
 }
@@ -632,7 +658,8 @@ static void __iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
 
 	size = PAGE_ALIGN(size);
 	/*
-	 * @cpu_addr will be one of 3 things depending on how it was allocated:
+	 * @cpu_addr will be one of 4 things depending on how it was allocated:
+	 * - A remapped array of pages for contiguous allocations.
 	 * - A remapped array of pages from iommu_dma_alloc(), for all
 	 *   non-atomic allocations.
 	 * - A non-cacheable alias from the atomic pool, for atomic
@@ -644,6 +671,12 @@ static void __iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
 	if (__in_atomic_pool(cpu_addr, size)) {
 		iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
 		__free_from_pool(cpu_addr, size);
+	} else if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
+		struct page *page = vmalloc_to_page(cpu_addr);
+
+		iommu_dma_unmap_page(dev, handle, iosize, 0, attrs);
+		dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
+		dma_common_free_remap(cpu_addr, size, VM_USERMAP);
 	} else if (is_vmalloc_addr(cpu_addr)){
 		struct vm_struct *area = find_vm_area(cpu_addr);
 
-- 
2.7.4