On Sun, 2013-07-28 at 16:01 -0700, Eric Dumazet wrote:

On Sun, 2013-07-28 at 21:22 +0100, Ben Hutchings wrote:

quoted

Since we know lengths > 4K work, perhaps it would be worth testing with
the fragment cache size reduced to 16K?  The driver would never
previously have used RX buffers crossing 16K boundaries, except if SLOB
was used (and that's an unlikely combination).

Sure, please note the following maths :

NET_SKB_PAD + 1536 + sizeof(struct skb_shared_info) = 1920

16384/1920 = 8

32768/1920 = 17

I don't think atl1c is used in any critical host (given it doesn't even
provide RX checksums and GRO ...), so I will provide a patch doing mere
page allocations.

Oh well, look at code around line 2530

        * The atl1c chip can DMA to 64-bit addresses, but it uses a single
         * shared register for the high 32 bits, so only a single, aligned,
         * 4 GB physical address range can be used at a time.
         *
         * Supporting 64-bit DMA on this hardware is more trouble than it's
         * worth.  It is far easier to limit to 32-bit DMA than update
         * various kernel subsystems to support the mechanics required by a
         * fixed-high-32-bit system.
         */
        if ((pci_set_dma_mask(pdev, DMA_BIT_MASK(32)) != 0) ||
            (pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)) != 0)) {
                dev_err(&pdev->dev, "No usable DMA configuration,aborting\n");
                goto err_dma;
        }

It looks like we have a winner !

This $@!? really needs DMA32 allocations.

Currently only tested on TX patch, it needs same care on RX

diff --git a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
index 786a874..e2ee962 100644
--- a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
+++ b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c

@@ -1660,7 +1660,8 @@ static int atl1c_alloc_rx_buffer(struct atl1c_adapter *adapter)
 	while (next_info->flags & ATL1C_BUFFER_FREE) {
 		rfd_desc = ATL1C_RFD_DESC(rfd_ring, rfd_next_to_use);
 
-		skb = netdev_alloc_skb(adapter->netdev, adapter->rx_buffer_len);
+		skb = __netdev_alloc_skb(adapter->netdev, adapter->rx_buffer_len,
+					 GFP_ATOMIC | GFP_DMA32);
 		if (unlikely(!skb)) {
 			if (netif_msg_rx_err(adapter))
 				dev_warn(&pdev->dev, "alloc rx buffer failed\n");