x86/pci: fix broken ISA DMA

Rene Herman reported:

> commit 8779f2f
>
> "x86: don't try to allocate from DMA zone at first"
>
> breaks all of ISA DMA. Or all of ALSA ISA DMA at least. All
> ISA soundcards are silent following that commit -- no error
> messages, everything appears fine, just silence.

That patch is buggy. We had an implicit assumption that
dev = NULL for ISA devices that require 24bit DMA.

The recent work on x86 dma_alloc_coherent() breaks the ISA DMA buffer
allocation, which is represented by "dev = NULL" and requires 24bit
DMA implicitly.

Bisected-by: Rene Herman <rene.herman@keyaccess.nl>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Tested-by: Rene Herman <rene.herman@keyaccess.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

arch/x86/kernel/pci-dma.c

@@ -385,11 +385,13 @@ dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	if (dma_alloc_from_coherent_mem(dev, size, dma_handle, &memory))
 		return memory;
 
-	if (!dev)
+	if (!dev) {
 		dev = &fallback_dev;
+		gfp |= GFP_DMA;
+	}
 	dma_mask = dev->coherent_dma_mask;
 	if (dma_mask == 0)
-		dma_mask = DMA_32BIT_MASK;
+		dma_mask = (gfp & GFP_DMA) ? DMA_24BIT_MASK : DMA_32BIT_MASK;
 
 	/* Device not DMA able */
 	if (dev->dma_mask == NULL)
@@ -403,7 +405,7 @@ dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	   larger than 16MB and in this case we have a chance of
 	   finding fitting memory in the next higher zone first. If
 	   not retry with true GFP_DMA. -AK */
-	if (dma_mask <= DMA_32BIT_MASK)
+	if (dma_mask <= DMA_32BIT_MASK && !(gfp & GFP_DMA))
 		gfp |= GFP_DMA32;
 #endif