diff options
author | Maciej Fijalkowski <maciej.fijalkowski@intel.com> | 2019-10-24 01:11:24 -0700 |
---|---|---|
committer | Jeff Kirsher <jeffrey.t.kirsher@intel.com> | 2019-11-04 13:09:59 -0800 |
commit | aaf27254fdf9054c00cf2a9c479a52a16204d768 (patch) | |
tree | 563ee5a13519e8d63903f26796c3dfd5c8cbe325 /drivers/net/ethernet/intel/ice | |
parent | 59bb08080557589aaf577a99d329ccea38b55c95 (diff) |
ice: add build_skb() support
Driver is now prepared for building the skb around the existing Rx
buffer, so introduce the ice_build_skb responsible for it. Make use of
XDP's data_meta as well.
I've observed around 30% less CPU consumption with build_skb Rx path, in
comparison to legacy Rx. What stands behind such result is the avoidance
of flow_dissector (which we were diving into via eth_get_headlen) and no
memcpy calls.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Diffstat (limited to 'drivers/net/ethernet/intel/ice')
-rw-r--r-- | drivers/net/ethernet/intel/ice/ice_txrx.c | 60 |
1 files changed, 59 insertions, 1 deletions
diff --git a/drivers/net/ethernet/intel/ice/ice_txrx.c b/drivers/net/ethernet/intel/ice/ice_txrx.c index bd1f73af468c..40a29b9d3034 100644 --- a/drivers/net/ethernet/intel/ice/ice_txrx.c +++ b/drivers/net/ethernet/intel/ice/ice_txrx.c @@ -791,6 +791,60 @@ ice_get_rx_buf(struct ice_ring *rx_ring, struct sk_buff **skb, } /** + * ice_build_skb - Build skb around an existing buffer + * @rx_ring: Rx descriptor ring to transact packets on + * @rx_buf: Rx buffer to pull data from + * @xdp: xdp_buff pointing to the data + * + * This function builds an skb around an existing Rx buffer, taking care + * to set up the skb correctly and avoid any memcpy overhead. + */ +static struct sk_buff * +ice_build_skb(struct ice_ring *rx_ring, struct ice_rx_buf *rx_buf, + struct xdp_buff *xdp) +{ + unsigned int metasize = xdp->data - xdp->data_meta; +#if (PAGE_SIZE < 8192) + unsigned int truesize = ice_rx_pg_size(rx_ring) / 2; +#else + unsigned int truesize = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) + + SKB_DATA_ALIGN(xdp->data_end - + xdp->data_hard_start); +#endif + struct sk_buff *skb; + + /* Prefetch first cache line of first page. If xdp->data_meta + * is unused, this points exactly as xdp->data, otherwise we + * likely have a consumer accessing first few bytes of meta + * data, and then actual data. + */ + prefetch(xdp->data_meta); +#if L1_CACHE_BYTES < 128 + prefetch((void *)(xdp->data + L1_CACHE_BYTES)); +#endif + /* build an skb around the page buffer */ + skb = build_skb(xdp->data_hard_start, truesize); + if (unlikely(!skb)) + return NULL; + + /* must to record Rx queue, otherwise OS features such as + * symmetric queue won't work + */ + skb_record_rx_queue(skb, rx_ring->q_index); + + /* update pointers within the skb to store the data */ + skb_reserve(skb, xdp->data - xdp->data_hard_start); + __skb_put(skb, xdp->data_end - xdp->data); + if (metasize) + skb_metadata_set(skb, metasize); + + /* buffer is used by skb, update page_offset */ + ice_rx_buf_adjust_pg_offset(rx_buf, truesize); + + return skb; +} + +/** * ice_construct_skb - Allocate skb and populate it * @rx_ring: Rx descriptor ring to transact packets on * @rx_buf: Rx buffer to pull data from @@ -996,12 +1050,14 @@ static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget) if (!size) { xdp.data = NULL; xdp.data_end = NULL; + xdp.data_hard_start = NULL; + xdp.data_meta = NULL; goto construct_skb; } xdp.data = page_address(rx_buf->page) + rx_buf->page_offset; xdp.data_hard_start = xdp.data - ice_rx_offset(rx_ring); - xdp_set_data_meta_invalid(&xdp); + xdp.data_meta = xdp.data; xdp.data_end = xdp.data + size; rcu_read_lock(); @@ -1038,6 +1094,8 @@ static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget) construct_skb: if (skb) ice_add_rx_frag(rx_ring, rx_buf, skb, size); + else if (ice_ring_uses_build_skb(rx_ring)) + skb = ice_build_skb(rx_ring, rx_buf, &xdp); else skb = ice_construct_skb(rx_ring, rx_buf, &xdp); |