[PATCH net-next] net: mvneta: avoid possible cache misses in mvneta_rx_swbm

[PATCH net-next] net: mvneta: avoid possible cache misses in mvneta_rx_swbm

[Lorenzo Bianconi]

Do not use rx_desc pointers if possible since rx descriptors are stored in
uncached memory and dereferencing rx_desc pointers generate extra loads.
This patch improves XDP_DROP performance of ~ 110Kpps (700Kpps vs 590Kpps)
on Marvell Espressobin

Analyzed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 drivers/net/ethernet/marvell/mvneta.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index bb485745b72a..d095718355d3 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -2236,19 +2236,22 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp,
 	enum dma_data_direction dma_dir;
 	struct skb_shared_info *sinfo;
 
-	if (rx_desc->data_size > MVNETA_MAX_RX_BUF_SIZE) {
+	if (*size > MVNETA_MAX_RX_BUF_SIZE) {
 		len = MVNETA_MAX_RX_BUF_SIZE;
 		data_len += len;
 	} else {
-		len = rx_desc->data_size;
+		len = *size;
 		data_len += len - ETH_FCS_LEN;
 	}
+	*size = *size - len;
 
 	dma_dir = page_pool_get_dma_dir(rxq->page_pool);
 	dma_sync_single_for_cpu(dev->dev.parent,
 				rx_desc->buf_phys_addr,
 				len, dma_dir);
 
+	rx_desc->buf_phys_addr = 0;
+
 	/* Prefetch header */
 	prefetch(data);
 
@@ -2259,9 +2262,6 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp,
 
 	sinfo = xdp_get_shared_info_from_buff(xdp);
 	sinfo->nr_frags = 0;
-
-	*size = rx_desc->data_size - len;
-	rx_desc->buf_phys_addr = 0;
 }
 
 static void
@@ -2375,7 +2375,7 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 
 			size = rx_desc->data_size;
 			frame_sz = size - ETH_FCS_LEN;
-			desc_status = rx_desc->status;
+			desc_status = rx_status;
 
 			mvneta_swbm_rx_frame(pp, rx_desc, rxq, &xdp_buf,
 					     &size, page);
-- 
2.26.2

Re: [PATCH net-next] net: mvneta: avoid possible cache misses in mvneta_rx_swbm

[David Miller]

From: Lorenzo Bianconi <lorenzo@kernel.org>
Date: Tue, 29 Sep 2020 23:58:57 +0200

> Do not use rx_desc pointers if possible since rx descriptors are stored in
> uncached memory and dereferencing rx_desc pointers generate extra loads.
> This patch improves XDP_DROP performance of ~ 110Kpps (700Kpps vs 590Kpps)
> on Marvell Espressobin
> 
> Analyzed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>

Applied.