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(-)
--
2.30.2
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
Signed-off-by: Aleksander Jan Bajkowski <[email protected]>
---
arch/mips/include/asm/mach-lantiq/xway/xway_dma.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/mips/include/asm/mach-lantiq/xway/xway_dma.h b/arch/mips/include/asm/mach-lantiq/xway/xway_dma.h
index 31ca9151b539..3dee15c61c8a 100644
--- a/arch/mips/include/asm/mach-lantiq/xway/xway_dma.h
+++ b/arch/mips/include/asm/mach-lantiq/xway/xway_dma.h
@@ -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 */
--
2.30.2
NAT 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
Signed-off-by: Aleksander Jan Bajkowski <[email protected]>
---
drivers/net/ethernet/lantiq_xrx200.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/lantiq_xrx200.c b/drivers/net/ethernet/lantiq_xrx200.c
index 503fb99c5b90..7cebf4a601bc 100644
--- a/drivers/net/ethernet/lantiq_xrx200.c
+++ b/drivers/net/ethernet/lantiq_xrx200.c
@@ -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);
--
2.30.2
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
Signed-off-by: Aleksander Jan Bajkowski <[email protected]>
---
drivers/net/ethernet/lantiq_xrx200.c | 56 ++++++++++++++++++----------
1 file changed, 36 insertions(+), 20 deletions(-)
diff --git a/drivers/net/ethernet/lantiq_xrx200.c b/drivers/net/ethernet/lantiq_xrx200.c
index 7cebf4a601bc..41d11137cde0 100644
--- a/drivers/net/ethernet/lantiq_xrx200.c
+++ b/drivers/net/ethernet/lantiq_xrx200.c
@@ -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 & LTQ_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 & LTQ_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);
--
2.30.2
On 1/4/22 07:11, Aleksander Jan Bajkowski wrote:
> 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
>
> Signed-off-by: Aleksander Jan Bajkowski <[email protected]>
> ---
Not sure why GRO is not yet implemented in this driver...
Hello:
This series was applied to netdev/net-next.git (master)
by Jakub Kicinski <[email protected]>:
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!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
Hi Enric,
On 1/4/22 18:42, Eric Dumazet wrote:
>
> On 1/4/22 07:11, Aleksander Jan Bajkowski wrote:
>> 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
>>
>> 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.