Re: [PATCH 12/21] dma-iommu: factor atomic pool allocations into helpers
From: Robin Murphy
Date: Tue Apr 23 2019 - 06:01:50 EST
On 19/04/2019 09:23, Christoph Hellwig wrote:
On Thu, Apr 18, 2019 at 07:15:00PM +0100, Robin Murphy wrote:
Still, I've worked in the vm_map_pages() stuff pending in MM and given them
the same treatment to finish the picture. Both x86_64_defconfig and
i386_defconfig do indeed compile and link fine as I expected, so I really
would like to understand the concern around #ifdefs better.
This looks generally fine to me. One thing I'd like to do is to
generally make use of the fact that __iommu_dma_get_pages returns NULL
for the force contigous case as that cleans up a few things. Also
for the !DMA_REMAP case we need to try the page allocator when
dma_alloc_from_contiguous does not return a page. What do you thing
of the following incremental diff? If that is fine with you I can
fold that in and add back in the remaining patches from my series
not obsoleted by your patches and resend.
Wouldn't this suffice? Since we also use alloc_pages() in the coherent
atomic case, the free path should already be able to deal with it.
Let me take a proper look at v3 and see how it all looks in context.
Robin.
----->8-----
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 1bc8d1de1a1d..0a02ddc27862 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -944,6 +944,8 @@ static void *iommu_dma_alloc(struct device *dev,
size_t size,
(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
page = dma_alloc_from_contiguous(dev, count, page_order,
gfp & __GFP_NOWARN);
+ if (!page)
+ page = alloc_pages(gfp, page_order);
} else {
return iommu_dma_alloc_remap(dev, size, handle, gfp, attrs);
}