Skip to contentSkip to conversation

dma-direct: fix the page free when it is not addressable

PATCH In review Chen Yu 2 participants 4 messages Aug 31, 2024 – Sep 3, 2024 DMA mapping helpers

lore
Share & tools
lore

Chen Yu

author
#0 lore raw

PATCH 1/1 dma-direct: fix the page free when it is not addressable

When the CMA allocation succeeds but isn't addressable, its buffer has already been released and the page is set to NULL. So later when the normal page allocation succeeds but isn't addressable, __free_pages() should be used to free that normal page rather than freeing continuous page(CMA).

Fixes 90ae409f9eb3 ("dma-direct: fix zone selection after an unaddressable CMA allocation")
---
 kernel/dma/direct.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/dma/direct.c +1 -1
index 4480a3cd92e0..51f07bf235c0 100644
@@ -140,7 +140,7 @@ static struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
 	if (!page)
 		page = alloc_pages_node(node, gfp, get_order(size));
 	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
-		dma_free_contiguous(dev, page, size);
+		__free_pages(page, get_order(size));
 		page = NULL;
 
 		if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Signature · 1 line
2.25.1
#1 lore raw
Quoted 5 lines · Chen Yu

When the CMA allocation succeeds but isn't addressable, its buffer has already been released and the page is set to NULL. So later when the normal page allocation succeeds but isn't addressable, __free_pages() should be used to free that normal page rather than freeing continuous page(CMA).

So the patch is obviously correct and probably useful, but I don't think the existing code is buggy.

dma_free_contiguous calls into cma_release, which uses cma_pages_valid to check if the page belongs to the CMA pool and retuns early if not, letting dma_free_contiguous fall back to __free_pages eventually.

What am I missing here?

author
#2 lore raw

Hi Christoph,

Quoted 7 lines · Christoph Hellwig
Quoted 5 lines · Chen Yu

When the CMA allocation succeeds but isn't addressable, its buffer has already been released and the page is set to NULL. So later when the normal page allocation succeeds but isn't addressable, __free_pages() should be used to free that normal page rather than freeing continuous page(CMA).

Quoted 9 lines

So the patch is obviously correct and probably useful, but I don't think the existing code is buggy.

dma_free_contiguous calls into cma_release, which uses cma_pages_valid to check if the page belongs to the CMA pool and retuns early if not, letting dma_free_contiguous fall back to __free_pages eventually.

What am I missing here?

Thanks for taking a look. Your are right, the pfn will be checked in cma_pages_valid() to see it is within the CMA range, if not, it prints some messages and return false, finally falls into __free_pages(). From the functional point of view, there is no bug. From the efficiency point of view(extra checks/printed message), and from the code readability to avoid confusing, maybe we can refine it?

thanks,
Chenyu

#3 lore raw
Quoted 6 lines · Chen Yu

Thanks for taking a look. Your are right, the pfn will be checked in cma_pages_valid() to see it is within the CMA range, if not, it prints some messages and return false, finally falls into __free_pages(). From the functional point of view, there is no bug. From the efficiency point of view(extra checks/printed message), and from the code readability to avoid confusing, maybe we can refine it?

I'll apply the patch with an updated commit log, thanks.