Ok, it seems like the sound noncontig alloc code that I already
commented on as potentially bogus GFP_GOMP mapping trips this. I think
for now the right thing would be to revert the hunk in dma-iommu.c
(see patch below). The other thing to try would be to remove both
uses GFP_COMP in sound/core/memalloc.c, which should have the same
effect.
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 9297b741f5e80e..f798c44e090337 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -744,9 +744,6 @@ static struct page **__iommu_dma_alloc_pages(struct device *dev,
/* IOMMU can map any pages, so himem can also be used here */
gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
- /* It makes no sense to muck about with huge pages */
- gfp &= ~__GFP_COMP;
-
while (count) {
struct page *page = NULL;
unsigned int order_size;
