Skip to contentSkip to conversation

bcm63xx_enet: major makeover of driver

PATCH Superseded Sieng Piaw Liew 3 participants 15 messages Dec 24, 2020 – Dec 29, 2020

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover bcm63xx_enet: major makeover of driver

This patch series aim to improve the bcm63xx_enet driver by integrating the latest networking features, i.e. batched rx processing, BQL, build_skb, etc.

The newer enetsw SoCs are found to be able to do unaligned rx DMA by adding NET_IP_ALIGN padding which, combined with these patches, improved packet processing performance by ~50% on BCM6328.

Older non-enetsw SoCs still benefit mainly from rx batching. Performance improvement of ~30% is observed on BCM6333.

The BCM63xx SoCs are designed for routers. As such, having BQL is beneficial as well as trivial to add.

v2:
* Add xmit_more support and rx loop improvisation patches.
* Moved BQL netdev_reset_queue() to bcm_enet_stop()/bcm_enetsw_stop() functions as suggested by Florian Fainelli.
* Improved commit messages.

Sieng Piaw Liew (6):
bcm63xx_enet: batch process rx path
bcm63xx_enet: add BQL support
bcm63xx_enet: add xmit_more support
bcm63xx_enet: alloc rx skb with NET_IP_ALIGN
bcm63xx_enet: convert to build_skb
bcm63xx_enet: improve rx loop

 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 184 ++++++++++---------
 drivers/net/ethernet/broadcom/bcm63xx_enet.h |  14 +-
 2 files changed, 103 insertions(+), 95 deletions(-)
Signature · 1 line
2.17.1
#1 lore raw

PATCH v2 1/6 bcm63xx_enet: batch process rx path

Use netif_receive_skb_list to batch process rx skb.
Tested on BCM6328 320 MHz using iperf3 -M 512, increasing performance by 12.5%.

Before:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   120 MBytes  33.7 Mbits/sec  277         sender
[  4]   0.00-30.00  sec   120 MBytes  33.5 Mbits/sec            receiver

After:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   136 MBytes  37.9 Mbits/sec  203         sender
[  4]   0.00-30.00  sec   135 MBytes  37.7 Mbits/sec            receiver
---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +5 -1
index 916824cca3fd..b82b7805c36a 100644
@@ -297,10 +297,12 @@ static void bcm_enet_refill_rx_timer(struct timer_list *t)
 static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 {
 	struct bcm_enet_priv *priv;
+	struct list_head rx_list;
 	struct device *kdev;
 	int processed;
 
 	priv = netdev_priv(dev);
+	INIT_LIST_HEAD(&rx_list);
 	kdev = &priv->pdev->dev;
 	processed = 0;
 
@@ -391,10 +393,12 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 		skb->protocol = eth_type_trans(skb, dev);
 		dev->stats.rx_packets++;
 		dev->stats.rx_bytes += len;
-		netif_receive_skb(skb);
+		list_add_tail(&skb->list, &rx_list);
 
 	} while (--budget > 0);
 
+	netif_receive_skb_list(&rx_list);
+
 	if (processed || !priv->rx_desc_count) {
 		bcm_enet_refill_rx(dev);
 
Signature · 1 line
2.17.1
Acked-by on v2 1/6
#2 lore raw

Re: [PATCH net-next v2 1/6] bcm63xx_enet: batch process rx path

Quoted 9 lines · Sieng Piaw Liew

Use netif_receive_skb_list to batch process rx skb.
Tested on BCM6328 320 MHz using iperf3 -M 512, increasing performance by 12.5%.

Before:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   120 MBytes  33.7 Mbits/sec  277         sender
[  4]   0.00-30.00  sec   120 MBytes  33.5 Mbits/sec            receiver
Quoted 6 lines

After:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   136 MBytes  37.9 Mbits/sec  203         sender
[  4]   0.00-30.00  sec   135 MBytes  37.7 Mbits/sec            receiver
Signature · 1 line
Florian
#3 lore raw

PATCH v2 2/6 bcm63xx_enet: add BQL support

In reply to Sieng Piaw Liew

Add Byte Queue Limits support to reduce/remove bufferbloat in bcm63xx_enet.

---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 9 +++++++++
 1 file changed, 9 insertions(+)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +9 -0
index b82b7805c36a..90f8214b4d22 100644
@@ -417,9 +417,11 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 static int bcm_enet_tx_reclaim(struct net_device *dev, int force)
 {
 	struct bcm_enet_priv *priv;
+	unsigned int bytes;
 	int released;
 
 	priv = netdev_priv(dev);
+	bytes = 0;
 	released = 0;
 
 	while (priv->tx_desc_count < priv->tx_ring_size) {
@@ -456,10 +458,13 @@ static int bcm_enet_tx_reclaim(struct net_device *dev, int force)
 		if (desc->len_stat & DMADESC_UNDER_MASK)
 			dev->stats.tx_errors++;
 
+		bytes += skb->len;
 		dev_kfree_skb(skb);
 		released++;
 	}
 
+	netdev_completed_queue(dev, released, bytes);
+
 	if (netif_queue_stopped(dev) && released)
 		netif_wake_queue(dev);
 
@@ -626,6 +631,8 @@ bcm_enet_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	desc->len_stat = len_stat;
 	wmb();
 
+	netdev_sent_queue(dev, skb->len);
+
 	/* kick tx dma */
 	enet_dmac_writel(priv, priv->dma_chan_en_mask,
 				 ENETDMAC_CHANCFG, priv->tx_chan);
@@ -1169,6 +1176,7 @@ static int bcm_enet_stop(struct net_device *dev)
 	kdev = &priv->pdev->dev;
 
 	netif_stop_queue(dev);
+	netdev_reset_queue(dev);
 	napi_disable(&priv->napi);
 	if (priv->has_phy)
 		phy_stop(dev->phydev);
@@ -2338,6 +2346,7 @@ static int bcm_enetsw_stop(struct net_device *dev)
 
 	del_timer_sync(&priv->swphy_poll);
 	netif_stop_queue(dev);
+	netdev_reset_queue(dev);
 	napi_disable(&priv->napi);
 	del_timer_sync(&priv->rx_timeout);
 
Signature · 1 line
2.17.1
#5 lore raw

PATCH v2 3/6 bcm63xx_enet: add xmit_more support

In reply to Sieng Piaw Liew

Support bulking hardware TX queue by using netdev_xmit_more().

---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +7 -4
index 90f8214b4d22..452968f168ed 100644
@@ -633,14 +633,17 @@ bcm_enet_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	netdev_sent_queue(dev, skb->len);
 
-	/* kick tx dma */
-	enet_dmac_writel(priv, priv->dma_chan_en_mask,
-				 ENETDMAC_CHANCFG, priv->tx_chan);
-
 	/* stop queue if no more desc available */
 	if (!priv->tx_desc_count)
 		netif_stop_queue(dev);
 
+	/* kick tx dma */
+        if (!netdev_xmit_more() || !priv->tx_desc_count)
+                enet_dmac_writel(priv, priv->dma_chan_en_mask,
+                                 ENETDMAC_CHANCFG, priv->tx_chan);
+
+
+
 	dev->stats.tx_bytes += skb->len;
 	dev->stats.tx_packets++;
 	ret = NETDEV_TX_OK;
Signature · 1 line
2.17.1
maintainer
#6 lore raw

Re: [PATCH net-next v2 3/6] bcm63xx_enet: add xmit_more support

Quoted 23 lines · On Thu, 24 Dec 2020 22:24:18 +0800 Sieng Piaw Liew wrote:

Support bulking hardware TX queue by using netdev_xmit_more().

---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +0 -4
index 90f8214b4d22..452968f168ed 100644
@@ -633,14 +633,17 @@ bcm_enet_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	netdev_sent_queue(dev, skb->len);
 
-	/* kick tx dma */
-	enet_dmac_writel(priv, priv->dma_chan_en_mask,
-				 ENETDMAC_CHANCFG, priv->tx_chan);
-
 	/* stop queue if no more desc available */
 	if (!priv->tx_desc_count)
 		netif_stop_queue(dev);
 

drivers/net/ethernet/broadcom/bcm63xx_enet.c bcm_enet_start_xmit(struct sk_buff *skb, struct net_device *dev) · line 649

Quoted 10 lines
drivers/net/ethernet/broadcom/bcm63xx_enet.c +7 -0
+	/* kick tx dma */
+        if (!netdev_xmit_more() || !priv->tx_desc_count)
+                enet_dmac_writel(priv, priv->dma_chan_en_mask,
+                                 ENETDMAC_CHANCFG, priv->tx_chan);
+
+
+
 	dev->stats.tx_bytes += skb->len;
 	dev->stats.tx_packets++;
 	ret = NETDEV_TX_OK;

Please address checkpatch --strict issues throughout the series:

ERROR: code indent should use tabs where possible #31: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:641:
+ if (!netdev_xmit_more() || !priv->tx_desc_count)$

WARNING: please, no spaces at the start of a line #31: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:641:
+ if (!netdev_xmit_more() || !priv->tx_desc_count)$

ERROR: code indent should use tabs where possible #32: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:642:
+ enet_dmac_writel(priv, priv->dma_chan_en_mask,$

WARNING: please, no spaces at the start of a line #32: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:642:
+ enet_dmac_writel(priv, priv->dma_chan_en_mask,$

ERROR: code indent should use tabs where possible #33: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:643:
+ ENETDMAC_CHANCFG, priv->tx_chan);$

WARNING: please, no spaces at the start of a line #33: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:643:
+ ENETDMAC_CHANCFG, priv->tx_chan);$

CHECK: Please don't use multiple blank lines
#35: FILE: drivers/net/ethernet/broadcom/bcm63xx_enet.c:645:
+
+

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#8 lore raw

PATCH v2 4/6 bcm63xx_enet: alloc rx skb with NET_IP_ALIGN

In reply to Sieng Piaw Liew

Use netdev_alloc_skb_ip_align on newer SoCs with integrated switch (enetsw) when refilling RX. Increases packet processing performance by 30% (with netif_receive_skb_list).

Non-enetsw SoCs cannot function with the extra pad so continue to use the regular netdev_alloc_skb.

Tested on BCM6328 320 MHz and iperf3 -M 512 to measure packet/sec performance.

Before:
[ ID] Interval Transfer Bandwidth Retr
[ 4] 0.00-30.00 sec 120 MBytes 33.7 Mbits/sec 277 sender [ 4] 0.00-30.00 sec 120 MBytes 33.5 Mbits/sec receiver

After (+netif_receive_skb_list):
[ 4] 0.00-30.00 sec 155 MBytes 43.3 Mbits/sec 354 sender [ 4] 0.00-30.00 sec 154 MBytes 43.1 Mbits/sec receiver

---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +4 -1
index 452968f168ed..51976ed87d2d 100644
@@ -237,7 +237,10 @@ static int bcm_enet_refill_rx(struct net_device *dev)
 		desc = &priv->rx_desc_cpu[desc_idx];
 
 		if (!priv->rx_skb[desc_idx]) {
-			skb = netdev_alloc_skb(dev, priv->rx_skb_size);
+			if (priv->enet_is_sw)
+				skb = netdev_alloc_skb_ip_align(dev, priv->rx_skb_size);
+			else
+				skb = netdev_alloc_skb(dev, priv->rx_skb_size);
 			if (!skb)
 				break;
 			priv->rx_skb[desc_idx] = skb;
Signature · 1 line
2.17.1
Acked-by on v2 4/6
#9 lore raw

Re: [PATCH net-next v2 4/6] bcm63xx_enet: alloc rx skb with NET_IP_ALIGN

Quoted 15 lines · Sieng Piaw Liew

Use netdev_alloc_skb_ip_align on newer SoCs with integrated switch (enetsw) when refilling RX. Increases packet processing performance by 30% (with netif_receive_skb_list).

Non-enetsw SoCs cannot function with the extra pad so continue to use the regular netdev_alloc_skb.

Tested on BCM6328 320 MHz and iperf3 -M 512 to measure packet/sec performance.

Before:
[ ID] Interval Transfer Bandwidth Retr
[ 4] 0.00-30.00 sec 120 MBytes 33.7 Mbits/sec 277 sender [ 4] 0.00-30.00 sec 120 MBytes 33.5 Mbits/sec receiver

Quoted 5 lines

After (+netif_receive_skb_list):
[ 4] 0.00-30.00 sec 155 MBytes 43.3 Mbits/sec 354 sender [ 4] 0.00-30.00 sec 154 MBytes 43.1 Mbits/sec receiver

Signature · 1 line
Florian
#10 lore raw

PATCH v2 5/6 bcm63xx_enet: convert to build_skb

In reply to Sieng Piaw Liew

We can increase the efficiency of rx path by using buffers to receive packets then build SKBs around them just before passing into the network stack. In contrast, preallocating SKBs too early reduces CPU cache efficiency.

Check if we're in NAPI context when refilling RX. Normally we're almost always running in NAPI context. Dispatch to napi_alloc_frag directly instead of relying on netdev_alloc_frag which does the same but with the overhead of local_bh_disable/enable.

Tested on BCM6328 320 MHz and iperf3 -M 512 to measure packet/sec performance. Included netif_receive_skb_list and NET_IP_ALIGN optimizations.

Before:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-10.00  sec  49.9 MBytes  41.9 Mbits/sec  197         sender
[  4]   0.00-10.00  sec  49.3 MBytes  41.3 Mbits/sec            receiver

After:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   171 MBytes  47.8 Mbits/sec  272         sender
[  4]   0.00-30.00  sec   170 MBytes  47.6 Mbits/sec            receiver
---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 157 +++++++++----------
 drivers/net/ethernet/broadcom/bcm63xx_enet.h |  14 +-
 2 files changed, 80 insertions(+), 91 deletions(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +70 -87
index 51976ed87d2d..8c2e97311a2c 100644
@@ -220,7 +220,7 @@ static void bcm_enet_mdio_write_mii(struct net_device *dev, int mii_id,
 /*
  * refill rx queue
  */
-static int bcm_enet_refill_rx(struct net_device *dev)
+static int bcm_enet_refill_rx(struct net_device *dev, bool napi_mode)
 {
 	struct bcm_enet_priv *priv;
 
@@ -228,29 +228,29 @@ static int bcm_enet_refill_rx(struct net_device *dev)
 
 	while (priv->rx_desc_count < priv->rx_ring_size) {
 		struct bcm_enet_desc *desc;
-		struct sk_buff *skb;
-		dma_addr_t p;
 		int desc_idx;
 		u32 len_stat;
 
 		desc_idx = priv->rx_dirty_desc;
 		desc = &priv->rx_desc_cpu[desc_idx];
 
-		if (!priv->rx_skb[desc_idx]) {
-			if (priv->enet_is_sw)
-				skb = netdev_alloc_skb_ip_align(dev, priv->rx_skb_size);
+		if (!priv->rx_buf[desc_idx]) {
+			void *buf;
+
+			if (likely(napi_mode))
+				buf = napi_alloc_frag(priv->rx_frag_size);
 			else
-				skb = netdev_alloc_skb(dev, priv->rx_skb_size);
-			if (!skb)
+				buf = netdev_alloc_frag(priv->rx_frag_size);
+			if (unlikely(!buf))
 				break;
-			priv->rx_skb[desc_idx] = skb;
-			p = dma_map_single(&priv->pdev->dev, skb->data,
-					   priv->rx_skb_size,
-					   DMA_FROM_DEVICE);
-			desc->address = p;
+			priv->rx_buf[desc_idx] = buf;
+			desc->address = dma_map_single(&priv->pdev->dev,
+						       buf + priv->rx_buf_offset,
+						       priv->rx_buf_size,
+						       DMA_FROM_DEVICE);
 		}
 
-		len_stat = priv->rx_skb_size << DMADESC_LENGTH_SHIFT;
+		len_stat = priv->rx_buf_size << DMADESC_LENGTH_SHIFT;
 		len_stat |= DMADESC_OWNER_MASK;
 		if (priv->rx_dirty_desc == priv->rx_ring_size - 1) {
 			len_stat |= (DMADESC_WRAP_MASK >> priv->dma_desc_shift);
@@ -290,7 +290,7 @@ static void bcm_enet_refill_rx_timer(struct timer_list *t)
 	struct net_device *dev = priv->net_dev;
 
 	spin_lock(&priv->rx_lock);
-	bcm_enet_refill_rx(dev);
+	bcm_enet_refill_rx(dev, false);
 	spin_unlock(&priv->rx_lock);
 }
 
@@ -320,6 +320,7 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 		int desc_idx;
 		u32 len_stat;
 		unsigned int len;
+		void *buf;
 
 		desc_idx = priv->rx_curr_desc;
 		desc = &priv->rx_desc_cpu[desc_idx];
@@ -365,16 +366,14 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 		}
 
 		/* valid packet */
-		skb = priv->rx_skb[desc_idx];
+		buf = priv->rx_buf[desc_idx];
 		len = (len_stat & DMADESC_LENGTH_MASK) >> DMADESC_LENGTH_SHIFT;
 		/* don't include FCS */
 		len -= 4;
 
 		if (len < copybreak) {
-			struct sk_buff *nskb;
-
-			nskb = napi_alloc_skb(&priv->napi, len);
-			if (!nskb) {
+			skb = napi_alloc_skb(&priv->napi, len);
+			if (unlikely(!skb)) {
 				/* forget packet, just rearm desc */
 				dev->stats.rx_dropped++;
 				continue;
@@ -382,14 +381,21 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 
 			dma_sync_single_for_cpu(kdev, desc->address,
 						len, DMA_FROM_DEVICE);
-			memcpy(nskb->data, skb->data, len);
+			memcpy(skb->data, buf + priv->rx_buf_offset, len);
 			dma_sync_single_for_device(kdev, desc->address,
 						   len, DMA_FROM_DEVICE);
-			skb = nskb;
 		} else {
-			dma_unmap_single(&priv->pdev->dev, desc->address,
-					 priv->rx_skb_size, DMA_FROM_DEVICE);
-			priv->rx_skb[desc_idx] = NULL;
+			dma_unmap_single(kdev, desc->address,
+					 priv->rx_buf_size, DMA_FROM_DEVICE);
+			priv->rx_buf[desc_idx] = NULL;
+
+			skb = build_skb(buf, priv->rx_frag_size);
+			if (unlikely(!skb)) {
+				skb_free_frag(buf);
+				dev->stats.rx_dropped++;
+				continue;
+			}
+			skb_reserve(skb, priv->rx_buf_offset);
 		}
 
 		skb_put(skb, len);
@@ -403,7 +409,7 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 	netif_receive_skb_list(&rx_list);
 
 	if (processed || !priv->rx_desc_count) {
-		bcm_enet_refill_rx(dev);
+		bcm_enet_refill_rx(dev, true);
 
 		/* kick rx dma */
 		enet_dmac_writel(priv, priv->dma_chan_en_mask,
@@ -862,6 +868,24 @@ static void bcm_enet_adjust_link(struct net_device *dev)
 		priv->pause_tx ? "tx" : "off");
 }
 
+static void bcm_enet_free_rx_buf_ring(struct device *kdev, struct bcm_enet_priv *priv)
+{
+	int i;
+
+	for (i = 0; i < priv->rx_ring_size; i++) {
+		struct bcm_enet_desc *desc;
+
+		if (!priv->rx_buf[i])
+			continue;
+
+		desc = &priv->rx_desc_cpu[i];
+		dma_unmap_single(kdev, desc->address, priv->rx_buf_size,
+				 DMA_FROM_DEVICE);
+		skb_free_frag(priv->rx_buf[i]);
+	}
+	kfree(priv->rx_buf);
+}
+
 /*
  * open callback, allocate dma rings & buffers and start rx operation
  */
@@ -971,10 +995,10 @@ static int bcm_enet_open(struct net_device *dev)
 	priv->tx_curr_desc = 0;
 	spin_lock_init(&priv->tx_lock);
 
-	/* init & fill rx ring with skbs */
-	priv->rx_skb = kcalloc(priv->rx_ring_size, sizeof(struct sk_buff *),
+	/* init & fill rx ring with buffers */
+	priv->rx_buf = kcalloc(priv->rx_ring_size, sizeof(void *),
 			       GFP_KERNEL);
-	if (!priv->rx_skb) {
+	if (!priv->rx_buf) {
 		ret = -ENOMEM;
 		goto out_free_tx_skb;
 	}
@@ -991,7 +1015,7 @@ static int bcm_enet_open(struct net_device *dev)
 		enet_dmac_writel(priv, ENETDMA_BUFALLOC_FORCE_MASK | 0,
 				ENETDMAC_BUFALLOC, priv->rx_chan);
 
-	if (bcm_enet_refill_rx(dev)) {
+	if (bcm_enet_refill_rx(dev, false)) {
 		dev_err(kdev, "cannot allocate rx skb queue\n");
 		ret = -ENOMEM;
 		goto out;
@@ -1086,18 +1110,7 @@ static int bcm_enet_open(struct net_device *dev)
 	return 0;
 
 out:
-	for (i = 0; i < priv->rx_ring_size; i++) {
-		struct bcm_enet_desc *desc;
-
-		if (!priv->rx_skb[i])
-			continue;
-
-		desc = &priv->rx_desc_cpu[i];
-		dma_unmap_single(kdev, desc->address, priv->rx_skb_size,
-				 DMA_FROM_DEVICE);
-		kfree_skb(priv->rx_skb[i]);
-	}
-	kfree(priv->rx_skb);
+	bcm_enet_free_rx_buf_ring(kdev, priv);
 
 out_free_tx_skb:
 	kfree(priv->tx_skb);
@@ -1176,7 +1189,6 @@ static int bcm_enet_stop(struct net_device *dev)
 {
 	struct bcm_enet_priv *priv;
 	struct device *kdev;
-	int i;
 
 	priv = netdev_priv(dev);
 	kdev = &priv->pdev->dev;
@@ -1204,21 +1216,10 @@ static int bcm_enet_stop(struct net_device *dev)
 	/* force reclaim of all tx buffers */
 	bcm_enet_tx_reclaim(dev, 1);
 
-	/* free the rx skb ring */
-	for (i = 0; i < priv->rx_ring_size; i++) {
-		struct bcm_enet_desc *desc;
-
-		if (!priv->rx_skb[i])
-			continue;
-
-		desc = &priv->rx_desc_cpu[i];
-		dma_unmap_single(kdev, desc->address, priv->rx_skb_size,
-				 DMA_FROM_DEVICE);
-		kfree_skb(priv->rx_skb[i]);
-	}
+	/* free the rx buffer ring */
+	bcm_enet_free_rx_buf_ring(kdev, priv);
 
 	/* free remaining allocated memory */
-	kfree(priv->rx_skb);
 	kfree(priv->tx_skb);
 	dma_free_coherent(kdev, priv->rx_desc_alloc_size,
 			  priv->rx_desc_cpu, priv->rx_desc_dma);
@@ -1640,9 +1641,12 @@ static int bcm_enet_change_mtu(struct net_device *dev, int new_mtu)
 	 * align rx buffer size to dma burst len, account FCS since
 	 * it's appended
 	 */
-	priv->rx_skb_size = ALIGN(actual_mtu + ETH_FCS_LEN,
+	priv->rx_buf_size = ALIGN(actual_mtu + ETH_FCS_LEN,
 				  priv->dma_maxburst * 4);
 
+	priv->rx_frag_size = SKB_DATA_ALIGN(priv->rx_buf_offset + priv->rx_buf_size)
+						+ SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+
 	dev->mtu = new_mtu;
 	return 0;
 }
@@ -1727,6 +1731,7 @@ static int bcm_enet_probe(struct platform_device *pdev)
 
 	priv->enet_is_sw = false;
 	priv->dma_maxburst = BCMENET_DMA_MAXBURST;
+	priv->rx_buf_offset = NET_SKB_PAD;
 
 	ret = bcm_enet_change_mtu(dev, dev->mtu);
 	if (ret)
@@ -2154,10 +2159,10 @@ static int bcm_enetsw_open(struct net_device *dev)
 	priv->tx_curr_desc = 0;
 	spin_lock_init(&priv->tx_lock);
 
-	/* init & fill rx ring with skbs */
-	priv->rx_skb = kcalloc(priv->rx_ring_size, sizeof(struct sk_buff *),
+	/* init & fill rx ring with buffers */
+	priv->rx_buf = kcalloc(priv->rx_ring_size, sizeof(void *),
 			       GFP_KERNEL);
-	if (!priv->rx_skb) {
+	if (!priv->rx_buf) {
 		dev_err(kdev, "cannot allocate rx skb queue\n");
 		ret = -ENOMEM;
 		goto out_free_tx_skb;
@@ -2205,7 +2210,7 @@ static int bcm_enetsw_open(struct net_device *dev)
 	enet_dma_writel(priv, ENETDMA_BUFALLOC_FORCE_MASK | 0,
 			ENETDMA_BUFALLOC_REG(priv->rx_chan));
 
-	if (bcm_enet_refill_rx(dev)) {
+	if (bcm_enet_refill_rx(dev, false)) {
 		dev_err(kdev, "cannot allocate rx skb queue\n");
 		ret = -ENOMEM;
 		goto out;
@@ -2305,18 +2310,7 @@ static int bcm_enetsw_open(struct net_device *dev)
 	return 0;
 
 out:
-	for (i = 0; i < priv->rx_ring_size; i++) {
-		struct bcm_enet_desc *desc;
-
-		if (!priv->rx_skb[i])
-			continue;
-
-		desc = &priv->rx_desc_cpu[i];
-		dma_unmap_single(kdev, desc->address, priv->rx_skb_size,
-				 DMA_FROM_DEVICE);
-		kfree_skb(priv->rx_skb[i]);
-	}
-	kfree(priv->rx_skb);
+	bcm_enet_free_rx_buf_ring(kdev, priv);
 
 out_free_tx_skb:
 	kfree(priv->tx_skb);
@@ -2345,7 +2339,6 @@ static int bcm_enetsw_stop(struct net_device *dev)
 {
 	struct bcm_enet_priv *priv;
 	struct device *kdev;
-	int i;
 
 	priv = netdev_priv(dev);
 	kdev = &priv->pdev->dev;
@@ -2367,21 +2360,10 @@ static int bcm_enetsw_stop(struct net_device *dev)
 	/* force reclaim of all tx buffers */
 	bcm_enet_tx_reclaim(dev, 1);
 
-	/* free the rx skb ring */
-	for (i = 0; i < priv->rx_ring_size; i++) {
-		struct bcm_enet_desc *desc;
-
-		if (!priv->rx_skb[i])
-			continue;
-
-		desc = &priv->rx_desc_cpu[i];
-		dma_unmap_single(kdev, desc->address, priv->rx_skb_size,
-				 DMA_FROM_DEVICE);
-		kfree_skb(priv->rx_skb[i]);
-	}
+	/* free the rx buffer ring */
+	bcm_enet_free_rx_buf_ring(kdev, priv);
 
 	/* free remaining allocated memory */
-	kfree(priv->rx_skb);
 	kfree(priv->tx_skb);
 	dma_free_coherent(kdev, priv->rx_desc_alloc_size,
 			  priv->rx_desc_cpu, priv->rx_desc_dma);
@@ -2678,6 +2660,7 @@ static int bcm_enetsw_probe(struct platform_device *pdev)
 	priv->rx_ring_size = BCMENET_DEF_RX_DESC;
 	priv->tx_ring_size = BCMENET_DEF_TX_DESC;
 	priv->dma_maxburst = BCMENETSW_DMA_MAXBURST;
+	priv->rx_buf_offset = NET_SKB_PAD + NET_IP_ALIGN;
 
 	pd = dev_get_platdata(&pdev->dev);
 	if (pd) {
drivers/net/ethernet/broadcom/bcm63xx_enet.h +10 -4
index 1d3c917eb830..78f1830fb3cb 100644
@@ -230,11 +230,17 @@ struct bcm_enet_priv {
 	/* next dirty rx descriptor to refill */
 	int rx_dirty_desc;
 
-	/* size of allocated rx skbs */
-	unsigned int rx_skb_size;
+	/* size of allocated rx buffers */
+	unsigned int rx_buf_size;
 
-	/* list of skb given to hw for rx */
-	struct sk_buff **rx_skb;
+	/* allocated rx buffer offset */
+	unsigned int rx_buf_offset;
+
+	/* size of allocated rx frag */
+	unsigned int rx_frag_size;
+
+	/* list of buffer given to hw for rx */
+	void **rx_buf;
 
 	/* used when rx skb allocation failed, so we defer rx queue
 	 * refill */
Signature · 1 line
2.17.1
#11 lore raw

Re: [PATCH net-next v2 5/6] bcm63xx_enet: convert to build_skb

Quoted 15 lines · Sieng Piaw Liew

We can increase the efficiency of rx path by using buffers to receive packets then build SKBs around them just before passing into the network stack. In contrast, preallocating SKBs too early reduces CPU cache efficiency.

Check if we're in NAPI context when refilling RX. Normally we're almost always running in NAPI context. Dispatch to napi_alloc_frag directly instead of relying on netdev_alloc_frag which does the same but with the overhead of local_bh_disable/enable.

Tested on BCM6328 320 MHz and iperf3 -M 512 to measure packet/sec performance. Included netif_receive_skb_list and NET_IP_ALIGN optimizations.

Before:

Quoted 8 lines
[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-10.00  sec  49.9 MBytes  41.9 Mbits/sec  197         sender
[  4]   0.00-10.00  sec  49.3 MBytes  41.3 Mbits/sec            receiver

After:

[ ID] Interval           Transfer     Bandwidth       Retr
[  4]   0.00-30.00  sec   171 MBytes  47.8 Mbits/sec  272         sender
[  4]   0.00-30.00  sec   170 MBytes  47.6 Mbits/sec            receiver

This looks good, however there are a few nits and suggestions below:

[snip]

Quoted 11 lines
drivers/net/ethernet/broadcom/bcm63xx_enet.c +7 -0
@@ -862,6 +868,24 @@ static void bcm_enet_adjust_link(struct net_device *dev)
 		priv->pause_tx ? "tx" : "off");
 }
 
+static void bcm_enet_free_rx_buf_ring(struct device *kdev, struct bcm_enet_priv *priv)
+{
+	int i;
+
+	for (i = 0; i < priv->rx_ring_size; i++) {
+		struct bcm_enet_desc *desc;
+

drivers/net/ethernet/broadcom/bcm63xx_enet.c static void bcm_enet_adjust_link(struct net_device *dev) · line 887

Quoted 10 lines
drivers/net/ethernet/broadcom/bcm63xx_enet.c +10 -0
+		if (!priv->rx_buf[i])
+			continue;
+
+		desc = &priv->rx_desc_cpu[i];
+		dma_unmap_single(kdev, desc->address, priv->rx_buf_size,
+				 DMA_FROM_DEVICE);
+		skb_free_frag(priv->rx_buf[i]);
+	}
+	kfree(priv->rx_buf);
+}

This is a good helper to introduced, however I would introduce it as a preliminary patch that way it becomes clear when you are doing the sk_buff to buf substitution in the next patch.

[snip]

drivers/net/ethernet/broadcom/bcm63xx_enet.c static int bcm_enet_change_mtu(struct net_device *dev, int new_mtu) · line 1648

Quoted 10 lines
drivers/net/ethernet/broadcom/bcm63xx_enet.c +3 -1
@@ -1640,9 +1641,12 @@ static int bcm_enet_change_mtu(struct net_device *dev, int new_mtu)
 	 * align rx buffer size to dma burst len, account FCS since
 	 * it's appended
 	 */
-	priv->rx_skb_size = ALIGN(actual_mtu + ETH_FCS_LEN,
+	priv->rx_buf_size = ALIGN(actual_mtu + ETH_FCS_LEN,
 				  priv->dma_maxburst * 4);
 
+	priv->rx_frag_size = SKB_DATA_ALIGN(priv->rx_buf_offset + priv->rx_buf_size)
+						+ SKB_DATA_ALIGN(sizeof(struct skb_shared_info));

The alignment of the second line is a bit off and you should aim for the
+ operator to be on the preceding line, and have SKB_DATA_ALIGN() start on the opening parenthesis of the preceding line.

Signature · 1 line
Florian
#12 lore raw

PATCH v2 6/6 bcm63xx_enet: improve rx loop

In reply to Sieng Piaw Liew

Use existing rx processed count to track against budget, thereby making budget decrement operation redundant.

rx_desc_count can be calculated outside the rx loop, making the loop a bit smaller.

---
 drivers/net/ethernet/broadcom/bcm63xx_enet.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/broadcom/bcm63xx_enet.c +2 -2
index 8c2e97311a2c..5ff0d39be2b2 100644
@@ -339,7 +339,6 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 		priv->rx_curr_desc++;
 		if (priv->rx_curr_desc == priv->rx_ring_size)
 			priv->rx_curr_desc = 0;
-		priv->rx_desc_count--;
 
 		/* if the packet does not have start of packet _and_
 		 * end of packet flag set, then just recycle it */
@@ -404,9 +403,10 @@ static int bcm_enet_receive_queue(struct net_device *dev, int budget)
 		dev->stats.rx_bytes += len;
 		list_add_tail(&skb->list, &rx_list);
 
-	} while (--budget > 0);
+	} while (processed < budget);
 
 	netif_receive_skb_list(&rx_list);
+	priv->rx_desc_count -= processed;
 
 	if (processed || !priv->rx_desc_count) {
 		bcm_enet_refill_rx(dev, true);
Signature · 1 line
2.17.1
maintainer
#14 lore raw
In reply to Sieng Piaw Liew
Quoted 7 lines · On Thu, 24 Dec 2020 22:24:15 +0800 Sieng Piaw Liew wrote:

This patch series aim to improve the bcm63xx_enet driver by integrating the latest networking features, i.e. batched rx processing, BQL, build_skb, etc.

The newer enetsw SoCs are found to be able to do unaligned rx DMA by adding NET_IP_ALIGN padding which, combined with these patches, improved packet processing performance by ~50% on BCM6328.

Quoted 5 lines

Older non-enetsw SoCs still benefit mainly from rx batching. Performance improvement of ~30% is observed on BCM6333.

The BCM63xx SoCs are designed for routers. As such, having BQL is beneficial as well as trivial to add.

Hopefully we can get some reviews now, but for inclusion in the tree you'll need to repost once net-next opens (should happen in the next few days):

http://vger.kernel.org/~davem/net-next.html

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel