The receive buffer offset includes NET_IP_ALIGN to get the IP header aligned behind a 14 byte Ethernet header. Most ag71xx devices use DSA switches whose tag sits in front of the EtherType, which moves the IP header by another 2 bytes. Every access to the IP and TCP headers then traps and gets emulated by the unaligned access handler, around 6 times per forwarded frame. Take the headroom needed by the DSA tagger into account when choosing the offset. TP-Link Archer C7 v2 (qca8k), median of 3 runs, Mbit/s: before after routed up 743 855 routed down 660 749 local receive 554 572 The unaligned instruction counter no longer moves while forwarding. Assisted-by: LLM Signed-off-by: Rosen Penev Reviewed-by: Andrew Lunn --- v2: rebase and ass Reviewed-by. drivers/net/ethernet/atheros/ag71xx.c | 24 ++++++++++++++++++++---- 1 file changed, 20 insertions(+), 4 deletions(-) diff --git a/drivers/net/ethernet/atheros/ag71xx.c b/drivers/net/ethernet/atheros/ag71xx.c index 4e4794c4dfdc..ff30ebce448d 100644 --- a/drivers/net/ethernet/atheros/ag71xx.c +++ b/drivers/net/ethernet/atheros/ag71xx.c @@ -38,6 +38,7 @@ #include #include #include +#include #include /* For our NAPI weight bigger does *NOT* mean better - it means more @@ -1401,6 +1402,24 @@ static void ag71xx_hw_disable(struct ag71xx *ag) ag71xx_rings_cleanup(ag); } +static unsigned int ag71xx_rx_offset(struct ag71xx *ag) +{ + unsigned int align = NET_IP_ALIGN; + + if (ag71xx_is(ag, AR7100) || ag71xx_is(ag, AR9130)) + return NET_SKB_PAD; + +#if IS_ENABLED(CONFIG_NET_DSA) + /* A tag inserted in front of the EtherType shifts the IP header, + * compensate for it to avoid unaligned accesses in the stack. + */ + if (netdev_uses_dsa(ag->ndev)) + align = (align - ag->ndev->dsa_ptr->tag_ops->needed_headroom) & 3; +#endif + + return NET_SKB_PAD + align; +} + static int ag71xx_open(struct net_device *ndev) { struct ag71xx *ag = netdev_priv(ndev); @@ -1415,6 +1434,7 @@ static int ag71xx_open(struct net_device *ndev) } max_frame_len = ag71xx_max_frame_len(ndev->mtu); + ag->rx_buf_offset = ag71xx_rx_offset(ag); ag->rx_buf_size = SKB_DATA_ALIGN(max_frame_len + NET_SKB_PAD + NET_IP_ALIGN); @@ -1879,10 +1899,6 @@ static int ag71xx_probe(struct platform_device *pdev) ndev->min_mtu = 68; ndev->max_mtu = dcfg->max_frame_len - ag71xx_max_frame_len(0); - ag->rx_buf_offset = NET_SKB_PAD; - if (!ag71xx_is(ag, AR7100) && !ag71xx_is(ag, AR9130)) - ag->rx_buf_offset += NET_IP_ALIGN; - if (ag71xx_is(ag, AR7100)) { ag->tx_ring.desc_split = AG71XX_TX_RING_SPLIT; tx_size *= AG71XX_TX_RING_DS_PER_PKT; -- 2.56.0