sfc: Fix efx_rx_buf_offset() for recycled pages

This bug fix is only for stable branches older than 3.10.  The bug was
fixed upstream by commit 2768935a4660 ('sfc: reuse pages to avoid DMA
mapping/unmapping costs'), but that change is totally unsuitable for
stable.

Commit b590ace09d51 ('sfc: Fix efx_rx_buf_offset() in the presence of
swiotlb') added an explicit page_offset member to struct
efx_rx_buffer, which must be set consistently with the u.page and
dma_addr fields.  However, it failed to add the necessary assignment
in efx_resurrect_rx_buffer().  It also did not correct the calculation
of efx_rx_buffer::dma_addr in efx_resurrect_rx_buffer(), which assumes
that DMA-mapping a page will result in a page-aligned DMA address
(exactly what swiotlb violates).

Add the assignment of efx_rx_buffer::page_offset and change the
calculation of dma_addr to make use of it.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Acked-by: David Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index 7b3c068a9f52f7ca16cbd370423e61ed01bcc394..62311ad398e8ca43b5710ecc1017ea7afc236f95 100644 (file)
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -311,8 +311,9 @@ static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
 
        index = rx_queue->added_count & rx_queue->ptr_mask;
        new_buf = efx_rx_buffer(rx_queue, index);
-       new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
        new_buf->u.page = rx_buf->u.page;
+       new_buf->page_offset = rx_buf->page_offset ^ (PAGE_SIZE >> 1);
+       new_buf->dma_addr = state->dma_addr + new_buf->page_offset;
        new_buf->len = rx_buf->len;
        new_buf->flags = EFX_RX_BUF_PAGE;
        ++rx_queue->added_count;