When receiving a packet with multiple fragments, hardware may still touch the first fragment until the entire packet has been received. The driver therefore keeps the first fragment mapped for DMA until end of packet has been asserted, and delays its dma_sync call until then.
The driver tries to fit multiple receive buffers on one page. When using 3K receive buffers (e.g. using Jumbo frames and legacy-rx is turned off/build_skb is being used) on an architecture with 4K pages, the driver allocates an order 1 compound page and uses one page per receive buffer. To determine the correct offset for a delayed DMA sync of the first fragment of a multi-fragment packet, the driver then cannot just use PAGE_MASK on the DMA address but has to construct a mask based on the actual size of the backing page.
Using PAGE_MASK in the 3K RX buffer/4K page architecture configuration will always sync the first page of a compound page. With the SWIOTLB enabled this can lead to corrupted packets (zeroed out first fragment, re-used garbage from another packet) and various consequences, such as slow/stalling data transfers and connection resets. For example, testing on a link with MTU exceeding 3058 bytes on a host with SWIOTLB enabled (e.g. "iommu=soft swiotlb=262144,force") TCP transfers quickly fizzle out without this patch.
Cc: stable@vger.kernel.org Fixes: 0c5661ecc5dd7 ("ixgbe: fix crash in build_skb Rx code path") Signed-off-by: Markus Boehme markubo@amazon.com --- drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c index c5ec17d19c59..507ef3471301 100644 --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c @@ -1825,7 +1825,8 @@ static void ixgbe_dma_sync_frag(struct ixgbe_ring *rx_ring, struct sk_buff *skb) { if (ring_uses_build_skb(rx_ring)) { - unsigned long offset = (unsigned long)(skb->data) & ~PAGE_MASK; + unsigned long mask = (unsigned long)ixgbe_rx_pg_size(rx_ring) - 1; + unsigned long offset = (unsigned long)(skb->data) & mask;
dma_sync_single_range_for_cpu(rx_ring->dev, IXGBE_CB(skb)->dma,