There is a fallback logic, so it's better to not use the OOM killer in the allocations. Signed-off-by: Andi Kleen <ak@suse.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
@@ -73,6 +73,9 @@ dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
if (dma_mask == 0)
dma_mask = 0xffffffff;
+ /* Don't invoke OOM killer */
+ gfp |= __GFP_NORETRY;
+
/* Kludge to make it bug-to-bug compatible with i386. i386
uses the normal dma_mask for alloc_coherent. */
dma_mask &= *dev->dma_mask;