dma-mapping: postpone cpu addr translation on mmap

Postpone calling virt_to_page() translation on memory locations not
guaranteed to be backed by a struct page.  Try first to map memory from
the device coherent memory pool, then perform translation if that fails.

On some architectures, specifically SH when configured with the SPARSEMEM
memory model, assuming a struct page is always assigned to a memory
address lead to unexpected hangs during the virtual to page address
translation. This patch fixes that specific issue but applies in the
general case too.
Suggested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo+renesas@jmondi.org>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

drivers/base/dma-mapping.c

@@ -226,7 +226,6 @@ int dma_common_mmap(struct device *dev, struct vm_area_struct *vma,
 #ifndef CONFIG_ARCH_NO_COHERENT_DMA_MMAP
 	unsigned long user_count = vma_pages(vma);
 	unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
-	unsigned long pfn = page_to_pfn(virt_to_page(cpu_addr));
 	unsigned long off = vma->vm_pgoff;
 
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
@@ -234,12 +233,11 @@ int dma_common_mmap(struct device *dev, struct vm_area_struct *vma,
 	if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
 		return ret;
 
-	if (off < count && user_count <= (count - off)) {
+	if (off < count && user_count <= (count - off))
 		ret = remap_pfn_range(vma, vma->vm_start,
-				      pfn + off,
+				      page_to_pfn(virt_to_page(cpu_addr)) + off,
 				      user_count << PAGE_SHIFT,
 				      vma->vm_page_prot);
-	}
 #endif	/* !CONFIG_ARCH_NO_COHERENT_DMA_MMAP */
 
 	return ret;