[openib-general] [PATCH] IB/ipoib: DMA alignment on ppc64

[Michael S. Tsirkin]

Unaligned DMA is slow ppc64 systems - that's why this architecture
overrides NET_IP_ALIGN.  IPoIB should take this into account and align
DMA on this platform.

Signed-off-by: Michael S. Tsirkin <mst at mellanox.co.il>

---

Roland, what do you think?
This comes from reading linux/skbuff.h and asm-powerpc/system.h - I
don't actually have a ppc64 to test the performance impact.

diff --git a/drivers/infiniband/ulp/ipoib/ipoib_ib.c b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
index 8bf5e9e..cec1cb4 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_ib.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
@@ -125,7 +125,7 @@ static int ipoib_alloc_rx_skb(struct net
 	struct sk_buff *skb;
 	dma_addr_t addr;
 
-	skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
+	skb = dev_alloc_skb(IPOIB_BUF_SIZE + (NET_IP_ALIGN ? 4 : 0));
 	if (!skb)
 		return -ENOMEM;
 
@@ -133,8 +133,12 @@ static int ipoib_alloc_rx_skb(struct net
 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 	 * header.  So we need 4 more bytes to get to 48 and align the
 	 * IP header to a multiple of 16.
+	 *
+	 * But don't do this on architectures that override NET_IP_ALIGN -
+	 * see comment in linux/skbuff.h for reasons why.
 	 */
-	skb_reserve(skb, 4);
+	if (NET_IP_ALIGN)
+		skb_reserve(skb, 4);
 
 	addr = dma_map_single(priv->ca->dma_device,
 			      skb->data, IPOIB_BUF_SIZE,

-- 
MST