[PATCH net-next V6 3/3] net/mlx5e: Align header copy to cache line for Striding RQ non-linear

[Tariq Toukan]

From: Dragos Tatulea <dtatulea@xxxxxxxxxx>

In Striding RQ non-linear mode, there is a memcpy to pull the
header from the first fragment into the linear part of the skb.
As the header length is not aligned, it can cause cache thrashing
from a Read-Modify-Write cycle for the remaining bytes of the
cache line.

This patch changes the memcopy length to be aligned to the cache line.
The DMA sync is also aligned to cache line size accordingly. Note that
the original DMA sync is done on the initial conservative headlen
which is min(MLX5E_RX_MAX_HEAD, cqe_bcnt).

To show the improvement, a test was run with an XDP_DROP program
processing 64B packets at 100% CPU utilization over a single queue at
9000 MTU:

|----------+----------+------|
| Before   | After    | Diff |
|----------+----------+------|
| 3.6 Mpps | 3.8 Mpps | 5%   |
|----------+----------+------|

(CX7 NIC on Intel Xeon Platinum 8580 system)

While small packets profit most from this improvement, large packets
are not negatively affected (no regressions).

Suggested-by: Tariq Toukan <tariqt@xxxxxxxxxx>
Signed-off-by: Dragos Tatulea <dtatulea@xxxxxxxxxx>
Signed-off-by: Tariq Toukan <tariqt@xxxxxxxxxx>
---
 drivers/net/ethernet/mellanox/mlx5/core/en_rx.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
index 301b33419207..e5963e1b5309 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
@@ -1973,7 +1973,7 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
 
 		addr = page_pool_get_dma_addr_netmem(head_page->netmem);
 		dma_sync_single_for_cpu(rq->pdev, addr + head_offset,
-					ALIGN(headlen, sizeof(long)),
+					ALIGN(headlen, cache_line_size()),
 					rq->buff.map_dir);
 
 		headlen = eth_get_headlen(rq->netdev, head_addr, headlen);
@@ -2086,7 +2086,7 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
 
 		/* copy header */
 		skb_copy_to_linear_data(skb, head_addr,
-					ALIGN(headlen, sizeof(long)));
+					ALIGN(headlen, cache_line_size()));
 
 		/* skb linear part was allocated with headlen and aligned to long */
 		skb->tail += headlen;
-- 
2.44.0