[PATCH] net/macb: Use non-coherent memory for rx buffers

On 11/23/2012 05:12 PM, Joachim Eastwood :

Hi Nicolas,

On 23 November 2012 14:50, Nicolas Ferre [off-list ref] wrote:

quoted

From: Havard Skinnemoen <redacted>

Allocate regular pages to use as backing for the RX ring and use the
DMA API to sync the caches. This should give a bit better performance
since it allows the CPU to do burst transfers from memory. It is also
a necessary step on the way to reduce the amount of copying done by
the driver.

Signed-off-by: Havard Skinnemoen <redacted>
[nicolas.ferre at atmel.com: adapt to newer kernel]
Signed-off-by: Nicolas Ferre <redacted>
---
 drivers/net/ethernet/cadence/macb.c | 206 +++++++++++++++++++++++-------------
 drivers/net/ethernet/cadence/macb.h |  20 +++-
 2 files changed, 148 insertions(+), 78 deletions(-)

<snip>

quoted

diff --git a/drivers/net/ethernet/cadence/macb.h b/drivers/net/ethernet/cadence/macb.h
index 570908b..74e68a3 100644
--- a/drivers/net/ethernet/cadence/macb.h
+++ b/drivers/net/ethernet/cadence/macb.h

@@ -453,6 +453,23 @@ struct macb_dma_desc {
 #define MACB_TX_USED_SIZE                      1

 /**
+ * struct macb_rx_page - data associated with a page used as RX buffers
+ * @page: Physical page used as storage for the buffers
+ * @phys: DMA address of the page
+ *
+ * Each page is used to provide %MACB_RX_BUFFERS_PER_PAGE RX buffers.
+ * The page gets an initial reference when it is inserted into the
+ * ring, and an additional reference each time it is passed up the
+ * stack as a fragment. When all the buffers have been used, we drop
+ * the initial reference and allocate a new page. Any additional
+ * references are dropped when the higher layers free the skb.
+ */
+struct macb_rx_page {
+       struct page             *page;
+       dma_addr_t              phys;
+};
+
+/**
  * struct macb_tx_skb - data about an skb which is being transmitted
  * @skb: skb currently being transmitted
  * @mapping: DMA address of the skb's data buffer

@@ -543,7 +560,7 @@ struct macb {

        unsigned int            rx_tail;
        struct macb_dma_desc    *rx_ring;
-       void                    *rx_buffers;
+       struct macb_rx_page     *rx_page;

        unsigned int            tx_head, tx_tail;
        struct macb_dma_desc    *tx_ring;

@@ -564,7 +581,6 @@ struct macb {

        dma_addr_t              rx_ring_dma;
        dma_addr_t              tx_ring_dma;
-       dma_addr_t              rx_buffers_dma;

        struct mii_bus          *mii_bus;
        struct phy_device       *phy_dev;
--

struct macb is shared between at91_ether and macb. Removing
rx_buffers_dma and rx_buffers will break compilation on at91_ether.

OMG, you are absolutely right.

So please either leave the two struct members alone, for now, or fix
up at91_ether at the same time.

Well, I do not plan to touch at91_ether driver for the moment, so I
certainly will keep the two struct members for now.

I will wait a little more feedback before sending a v2 patch with these
changes.

Best regards,
-- 
Nicolas Ferre