Thread (48 messages) 48 messages, 4 authors, 2013-10-09

[PATCH v6 12/19] swiotlb: don't assume that io_tlb_start-io_tlb_end is coherent

From: Stefano Stabellini <hidden>
Date: 2013-10-02 17:32:43
Also in: lkml, xen-devel

On Mon, 30 Sep 2013, Konrad Rzeszutek Wilk wrote:
On Fri, Sep 27, 2013 at 05:10:00PM +0100, Stefano Stabellini wrote:
quoted
The swiotlb code has appropriate calls to dma_mark_clean in place for
buffers passed to swiotlb_map_page as an argument. However it assumes
that the swiotlb bounce buffer (io_tlb_start-io_tlb_end) is already
coherent and doesn't need any calls to dma_mark_clean.

On ARM the swiotlb bounce buffer is not coherent (the memory is
writealloc while it should be bufferable) and therefore we need to call
dma_mark_clean appropriately on the bounce buffer code paths too.

Note that most architecures have an empty dma_mark_clean implementation
anyway.
The other architecture that uses swiotlb is IA64 and that does have
an implementation where it touches on page attributes.

Which means I have to figure out why my HP zx6000 won't boot with 3.11 now :-(
Now this is a very thorny issue.

Honestly I don't like the dma_mark_clean interface very much: it's one
big hammer, when we actually need some finesse to handle coherency.

For example on ARM some devices might not need the dma_mark_clean call,
while others do. Calling it all the times is at the very best
inefficient and incorrect at worst.

I am thinking of calling the original map/unmap_page functions instead
(arm_dma_map_page or arm_coherent_dma_map_page in the arm case).
However in order to do that I would need to add more __get_dma_ops calls in
both lib/swiotlb.c and drivers/xen/swiotlb-xen.c

quoted
Signed-off-by: Stefano Stabellini <redacted>
---
 lib/swiotlb.c |   13 ++++++++++---
 1 files changed, 10 insertions(+), 3 deletions(-)
diff --git a/lib/swiotlb.c b/lib/swiotlb.c
index 4e8686c..eb45d17 100644
--- a/lib/swiotlb.c
+++ b/lib/swiotlb.c
@@ -515,6 +515,7 @@ found:
 		io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
 	if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
 		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
+	dma_mark_clean(phys_to_virt(tlb_addr), size);
 
 	return tlb_addr;
 }
@@ -547,7 +548,10 @@ void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
 	 * First, sync the memory before unmapping the entry
 	 */
 	if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
+	{
+		dma_mark_clean(phys_to_virt(tlb_addr), size);
 		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
+	}
 
 	/*
 	 * Return the buffer to the free list by setting the corresponding
@@ -587,17 +591,20 @@ void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
 
 	switch (target) {
 	case SYNC_FOR_CPU:
-		if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
+		if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)) {
+			dma_mark_clean(phys_to_virt(tlb_addr), size);
 			swiotlb_bounce(orig_addr, tlb_addr,
 				       size, DMA_FROM_DEVICE);
+		}
 		else
 			BUG_ON(dir != DMA_TO_DEVICE);
 		break;
 	case SYNC_FOR_DEVICE:
-		if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
+		if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) {
 			swiotlb_bounce(orig_addr, tlb_addr,
 				       size, DMA_TO_DEVICE);
-		else
+			dma_mark_clean(phys_to_virt(tlb_addr), size);
+		} else
 			BUG_ON(dir != DMA_FROM_DEVICE);
 		break;
 	default:
-- 
1.7.2.5
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help