net: lantiq_xrx200: improve ethernet performance
cover net: lantiq_xrx200: improve ethernet performance
This patchset improves Ethernet performance by 15%.
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 539 Mbps 599 Mbps
After 624 Mbps 695 Mbps
Aleksander Jan Bajkowski (3):
MIPS: lantiq: dma: increase descritor count
net: lantiq_xrx200: increase napi poll weigth
net: lantiq_xrx200: convert to build_skb
.../include/asm/mach-lantiq/xway/xway_dma.h | 2 +- drivers/net/ethernet/lantiq_xrx200.c | 62 ++++++++++++------- 2 files changed, 41 insertions(+), 23 deletions(-)
Signature · 1 line
2.30.2
PATCH 1/3 MIPS: lantiq: dma: increase descritor count
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 539 Mbps 599 Mbps
After 545 Mbps 625 Mbps
--- arch/mips/include/asm/mach-lantiq/xway/xway_dma.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/mips/include/asm/mach-lantiq/xway/xway_dma.h +1 -1
@@ -8,7 +8,7 @@#define LTQ_DMA_H__ #define LTQ_DESC_SIZE 0x08/* each descriptor is 64bit */ -#define LTQ_DESC_NUM 0x40/* 64 descriptors / channel */ +#define LTQ_DESC_NUM 0xC0/* 192 descriptors / channel */ #define LTQ_DMA_OWN BIT(31)/* owner bit */ #define LTQ_DMA_C BIT(30)/* complete bit */
Signature · 1 line
2.30.2
PATCH 2/3 net: lantiq_xrx200: increase napi poll weigth
In reply to Aleksander Jan BajkowskiNAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 545 Mbps 625 Mbps
After 577 Mbps 648 Mbps
--- drivers/net/ethernet/lantiq_xrx200.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/net/ethernet/lantiq_xrx200.c +4 -2
@@ -597,8 +597,10 @@ static int xrx200_probe(struct platform_device *pdev) PMAC_HD_CTL);/* setup NAPI */ -netif_napi_add (net_dev, &priv->chan_rx.napi, xrx200_poll_rx,32 ); -netif_tx_napi_add (net_dev, &priv->chan_tx.napi, xrx200_tx_housekeeping,32 ); +netif_napi_add (net_dev, &priv->chan_rx.napi, xrx200_poll_rx, +NAPI_POLL_WEIGHT ); +netif_tx_napi_add (net_dev, &priv->chan_tx.napi, xrx200_tx_housekeeping, +NAPI_POLL_WEIGHT );platform_set_drvdata (pdev, priv);
Signature · 1 line
2.30.2
PATCH 3/3 net: lantiq_xrx200: convert to build_skb
In reply to Aleksander Jan BajkowskiWe 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.
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 577 Mbps 648 Mbps
After 624 Mbps 695 Mbps
--- drivers/net/ethernet/lantiq_xrx200.c | 56 ++++++++++++++++++---------- 1 file changed, 36 insertions(+), 20 deletions(-)
drivers/net/ethernet/lantiq_xrx200.c +36 -20
@@ -63,7 +63,11 @@ struct xrx200_chan {struct napi_struct napi;struct ltq_dma_channel dma; -struct sk_buff *skb[LTQ_DESC_NUM ]; + +union { +struct sk_buff *skb[LTQ_DESC_NUM ]; +void *rx_buff[LTQ_DESC_NUM ]; + };struct sk_buff *skb_head;struct sk_buff *skb_tail;
@@ -78,6 +82,7 @@ struct xrx200_priv {struct xrx200_chan chan_rx;u16 rx_buf_size; +u16 rx_skb_size;struct net_device *net_dev;struct device *dev;
@@ -115,6 +120,12 @@ static int xrx200_buffer_size(int mtu)return round_up (xrx200_max_frame_len (mtu),4 *XRX200_DMA_BURST_LEN ); } +static int xrx200_skb_size (u16 buf_size) +{ +return SKB_DATA_ALIGN (buf_size +NET_SKB_PAD +NET_IP_ALIGN ) + +SKB_DATA_ALIGN (sizeof (struct skb_shared_info )); +} +/* drop all the packets from the DMA ring */ static void xrx200_flush_dma (struct xrx200_chan *ch) {
@@ -173,30 +184,29 @@ static int xrx200_close(struct net_device *net_dev)return 0 ; } -static int xrx200_alloc_skb (struct xrx200_chan *ch) +static int xrx200_alloc_buf (struct xrx200_chan *ch,void *(*alloc)(unsigned int size)) { -struct sk_buff *skb = ch->skb[ch->dma.desc]; +void *buf = ch->rx_buff[ch->dma.desc];struct xrx200_priv *priv = ch->priv;dma_addr_t mapping;int ret =0 ; - ch->skb[ch->dma.desc] =netdev_alloc_skb_ip_align (priv->net_dev, - priv->rx_buf_size); -if (!ch->skb[ch->dma.desc]) { + ch->rx_buff[ch->dma.desc] =alloc (priv->rx_skb_size); +if (!ch->rx_buff[ch->dma.desc]) { ret = -ENOMEM ; goto skip; } - mapping =dma_map_single (priv->dev, ch->skb[ch->dma.desc]->data, + mapping =dma_map_single (priv->dev, ch->rx_buff[ch->dma.desc], priv->rx_buf_size,DMA_FROM_DEVICE );if (unlikely (dma_mapping_error (priv->dev, mapping))) { -dev_kfree_skb_any (ch->skb[ch->dma.desc]); - ch->skb[ch->dma.desc] = skb; +skb_free_frag (ch->rx_buff[ch->dma.desc]); + ch->rx_buff[ch->dma.desc] = buf; ret = -ENOMEM ; goto skip; } - ch->dma.desc_base[ch->dma.desc].addr = mapping; + ch->dma.desc_base[ch->dma.desc].addr = mapping +NET_SKB_PAD +NET_IP_ALIGN ;/* Make sure the address is written before we give it to HW */ wmb (); skip:
@@ -210,13 +220,14 @@ static int xrx200_hw_receive(struct xrx200_chan *ch) {struct xrx200_priv *priv = ch->priv;struct ltq_dma_desc *desc = &ch->dma.desc_base[ch->dma.desc]; -struct sk_buff *skb = ch->skb[ch->dma.desc]; +void *buf = ch->rx_buff[ch->dma.desc];u32 ctl = desc->ctl;int len = (ctl <Q_DMA_SIZE_MASK );struct net_device *net_dev = priv->net_dev; +struct sk_buff *skb;int ret; - ret =xrx200_alloc_skb (ch); + ret =xrx200_alloc_buf (ch, napi_alloc_frag); ch->dma.desc++; ch->dma.desc %=LTQ_DESC_NUM ;
@@ -227,19 +238,21 @@ static int xrx200_hw_receive(struct xrx200_chan *ch)return ret; } + skb =build_skb (buf, priv->rx_skb_size); +skb_reserve (skb,NET_SKB_PAD );skb_put (skb, len);/* add buffers to skb via skb->frag_list */ if (ctl <Q_DMA_SOP ) { ch->skb_head = skb; ch->skb_tail = skb; +skb_reserve (skb,NET_IP_ALIGN ); }else if (ch->skb_head) {if (ch->skb_head == ch->skb_tail)skb_shinfo (ch->skb_tail)->frag_list = skb;else ch->skb_tail->next = skb; ch->skb_tail = skb; -skb_reserve (ch->skb_tail, -NET_IP_ALIGN ); ch->skb_head->len += skb->len; ch->skb_head->data_len += skb->len; ch->skb_head->truesize += skb->truesize;
@@ -395,12 +408,13 @@ xrx200_change_mtu(struct net_device *net_dev, int new_mtu)struct xrx200_chan *ch_rx = &priv->chan_rx;int old_mtu = net_dev->mtu;bool running = false; -struct sk_buff *skb; +void *buff;int curr_desc;int ret =0 ; net_dev->mtu = new_mtu; priv->rx_buf_size =xrx200_buffer_size (new_mtu); + priv->rx_skb_size =xrx200_skb_size (priv->rx_buf_size);if (new_mtu <= old_mtu)return ret;
@@ -416,14 +430,15 @@ xrx200_change_mtu(struct net_device *net_dev, int new_mtu)for (ch_rx->dma.desc =0 ; ch_rx->dma.desc <LTQ_DESC_NUM ; ch_rx->dma.desc++) { - skb = ch_rx->skb[ch_rx->dma.desc]; - ret =xrx200_alloc_skb (ch_rx); + buff = ch_rx->rx_buff[ch_rx->dma.desc]; + ret =xrx200_alloc_buf (ch_rx, netdev_alloc_frag);if (ret) { net_dev->mtu = old_mtu; priv->rx_buf_size =xrx200_buffer_size (old_mtu); + priv->rx_skb_size =xrx200_skb_size (priv->rx_buf_size);break ; } -dev_kfree_skb_any (skb); +skb_free_frag (buff); } ch_rx->dma.desc = curr_desc;
@@ -476,7 +491,7 @@ static int xrx200_dma_init(struct xrx200_priv *priv)ltq_dma_alloc_rx (&ch_rx->dma);for (ch_rx->dma.desc =0 ; ch_rx->dma.desc <LTQ_DESC_NUM ; ch_rx->dma.desc++) { - ret =xrx200_alloc_skb (ch_rx); + ret =xrx200_alloc_buf (ch_rx, netdev_alloc_frag);if (ret) goto rx_free; }
@@ -511,7 +526,7 @@ static int xrx200_dma_init(struct xrx200_priv *priv)/* free the allocated RX ring */ for (i =0 ; i <LTQ_DESC_NUM ; i++) {if (priv->chan_rx.skb[i]) -dev_kfree_skb_any (priv->chan_rx.skb[i]); +skb_free_frag (priv->chan_rx.rx_buff[i]); } rx_free:
@@ -528,7 +543,7 @@ static void xrx200_hw_cleanup(struct xrx200_priv *priv)/* free the allocated RX ring */ for (i =0 ; i <LTQ_DESC_NUM ; i++) -dev_kfree_skb_any (priv->chan_rx.skb[i]); +skb_free_frag (priv->chan_rx.rx_buff[i]); }static int xrx200_probe (struct platform_device *pdev)
@@ -553,6 +568,7 @@ static int xrx200_probe(struct platform_device *pdev) net_dev->min_mtu =ETH_ZLEN ; net_dev->max_mtu =XRX200_DMA_DATA_LEN -xrx200_max_frame_len (0 ); priv->rx_buf_size =xrx200_buffer_size (ETH_DATA_LEN ); + priv->rx_skb_size =xrx200_skb_size (priv->rx_buf_size);/* load the memory ranges */ priv->pmac_reg =devm_platform_get_and_ioremap_resource (pdev,0 ,NULL );
Signature · 1 line
2.30.2
Re: [PATCH net-next 3/3] net: lantiq_xrx200: convert to build_skb
Quoted 5 lines · Aleksander Jan Bajkowski
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.
Quoted 8 lines
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 577 Mbps 648 Mbps
After 624 Mbps 695 MbpsSigned-off-by Aleksander Jan Bajkowski <[email protected]>---
Not sure why GRO is not yet implemented in this driver...
Re: [PATCH net-next 3/3] net: lantiq_xrx200: convert to build_skb
Hi Enric,
Quoted 13 lines · Eric Dumazet
Quoted 11 lines · Aleksander Jan Bajkowski
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.
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 577 Mbps 648 Mbps
After 624 Mbps 695 Mbps
Quoted 7 lines
Quoted 2 lines
Signed-off-by Aleksander Jan Bajkowski <[email protected]>---
Not sure why GRO is not yet implemented in this driver...
I changed netif_receive_skb() to napi_gro_receive() and did some tests. I don't see any performance difference. Are any further changes needed to support GRO?
Maybe this is because this MAC is connected to a DSA switch and the napi_gro_receive() function cannot aggregate packets until the DSA tag is removed.
This series was applied to netdev/net-next.git (master) by Jakub Kicinski <[email protected]>:
Hello:
This series was applied to netdev/net-next.git (master) by Jakub Kicinski <[email protected]>:
Quoted 9 lines · On Tue, 4 Jan 2022 16:11:41 +0100 you wrote:
This patchset improves Ethernet performance by 15%.
NAT Performance results on BT Home Hub 5A (kernel 5.10.89, mtu 1500):
Down Up
Before 539 Mbps 599 Mbps
After 624 Mbps 695 Mbps[...]
Here is the summary with links:
- [net-next,1/3] MIPS: lantiq: dma: increase descritor count
https://git.kernel.org/netdev/net-next/c/5112e9234bbb - [net-next,2/3] net: lantiq_xrx200: increase napi poll weigth https://git.kernel.org/netdev/net-next/c/768818d772d5 - [net-next,3/3] net: lantiq_xrx200: convert to build_skb https://git.kernel.org/netdev/net-next/c/e015593573b3
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html