Thread (24 messages) 24 messages, 4 authors, 2011-08-18

[PATCH 8/9] ARM: integrate CMA with DMA-mapping subsystem

From: arnd@arndb.de (Arnd Bergmann)
Date: 2011-08-12 15:00:47
Also in: linux-media, linux-mm, lkml

On Friday 12 August 2011, Marek Szyprowski wrote:
quoted hunk ↗ jump to hunk
@@ -82,16 +103,16 @@ static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gf
 	if (mask < 0xffffffffULL)
 		gfp |= GFP_DMA;
 
-	page = alloc_pages(gfp, order);
-	if (!page)
-		return NULL;
-
 	/*
-	 * Now split the huge page and free the excess pages
+	 * Allocate contiguous memory
 	 */
-	split_page(page, order);
-	for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
-		__free_page(p);
+	if (cma_available())
+		page = dma_alloc_from_contiguous(dev, count, order);
+	else
+		page = __dma_alloc_system_pages(count, gfp, order);
+
+	if (!page)
+		return NULL;
Why do you need the fallback here? I would assume that CMA now has to be available
on ARMv6 and up to work at all. When you allocate from __dma_alloc_system_pages(),
wouldn't that necessarily fail in the dma_remap_area() stage?
 
-	if (arch_is_coherent() || nommu()) {
+	if (arch_is_coherent() || nommu() ||
+	   (cma_available() && !(gfp & GFP_ATOMIC))) {
+		/*
+		 * Allocate from system or CMA pages
+		 */
 		struct page *page = __dma_alloc_buffer(dev, size, gfp);
 		if (!page)
 			return NULL;
+		dma_remap_area(page, size, area->prot);
 		pfn = page_to_pfn(page);
 		ret = page_address(page);
Similarly with coherent and nommu. It seems to me that lumping too
many cases together creates extra complexity here.

How about something like

	if (arch_is_coherent() || nommu())
		ret = alloc_simple_buffer();
	else if (arch_is_v4_v5())
		ret = alloc_remap();
	else if (gfp & GFP_ATOMIC)
		ret = alloc_from_pool();
	else
		ret = alloc_from_contiguous();

This also allows a natural conversion to dma_map_ops when we get there.
 	/* reserve any platform specific memblock areas */
 	if (mdesc->reserve)
 		mdesc->reserve();
 
+	dma_coherent_reserve();
+	dma_contiguous_reserve();
+
 	memblock_analyze();
 	memblock_dump_all();
 }
Since we can handle most allocations using CMA on ARMv6+, I would think
that we can have a much smaller reserved area. Have you tried changing
dma_coherent_reserve() to allocate out of the contiguous area instead of
wasting a full 2MB section of memory?

	Arnd
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help