How can we make the lantiq xrx200 devices faster

nicefile · September 7, 2021, 12:22pm

I have good results using FB7412 and Open master 5.4 with pull 4353 and 4339 and with VPE-IRQ patch
Just three clicks to include them in OpenWrt master branch to benefit all

wilsonyan · October 20, 2021, 4:41am

now that there has been more of a push towards the kernel 5.10 and the DSA driver instead of the legacy driver it appers that there's been a fair amount of work to bring the things pc2005 worked on into the new driver

Extra stuff appears to be in this developer's branch : https://github.com/abajk/openwrt/commits/lantiq-jumbo-test

Like:

github.com

abajk/openwrt/blob/574c6559ebc499debc5cc11e093c7172871f7ed1/target/linux/lantiq/patches-5.10/0707-mips-lantiq-dma-test-packet-burst-arbitration.patch

From 37b7e58ee0e47a95d0fd3a1920de35c4e1fe9c74 Mon Sep 17 00:00:00 2001
From: Aleksander Jan Bajkowski <olek2@wp.pl>
Date: Fri, 8 Oct 2021 00:54:54 +0200
Subject: [PATCH] mips: lantiq: dma: test packet burst arbitration

---
 arch/mips/lantiq/xway/dma.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/mips/lantiq/xway/dma.c b/arch/mips/lantiq/xway/dma.c
index f8eedeb15f18..dbf576f12dc8 100644
--- a/arch/mips/lantiq/xway/dma.c
+++ b/arch/mips/lantiq/xway/dma.c
@@ -38,6 +38,7 @@
 #define DMA_PDEN		BIT(6)		/* enable packet drop */
 #define DMA_CHAN_RST		BIT(1)		/* channel on / off bit */
 #define DMA_RESET		BIT(0)		/* channel on / off bit */
+#define DMA_PKTARB		BIT(31)		/* packet arbitration */
 #define DMA_IRQ_ACK		0x7e		/* IRQ status register */
 #define DMA_POLL		BIT(31)		/* turn on channel polling */

This file has been truncated. show original

github.com

abajk/openwrt/blob/574c6559ebc499debc5cc11e093c7172871f7ed1/target/linux/lantiq/patches-5.10/0704-net-lantiq_xrx200-receive-fragmented-packets-via-fra.patch

From 17faec96dd31bfb338cc35200e4876965cfb915b Mon Sep 17 00:00:00 2001
From: Aleksander Jan Bajkowski <olek2@wp.pl>
Date: Sat, 9 Oct 2021 19:07:41 +0200
Subject: [PATCH] net: lantiq_xrx200: receive fragmented packets via frag_list

DMA split receive packet into multiple buffers in some cases.
If it occur create fragmented packet via frag_list.

Signed-off-by: Aleksander Jan Bajkowski <olek2@wp.pl>
---
 drivers/net/ethernet/lantiq_xrx200.c | 39 +++++++++++++++++++++++-----
 1 file changed, 33 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/lantiq_xrx200.c b/drivers/net/ethernet/lantiq_xrx200.c
index 9b7307eba97c..b6752fe460b0 100644
--- a/drivers/net/ethernet/lantiq_xrx200.c
+++ b/drivers/net/ethernet/lantiq_xrx200.c
@@ -26,6 +26,9 @@
 #define XRX200_DMA_RX		0
 #define XRX200_DMA_TX		1

This file has been truncated. show original

However I am unable to get the ethernet ports working on my BT home hub 5a after merging this branch. Can anyone with serial access (so they don't brick their router unless they have a pre-existing config set to bridge a usb ethernet adapter onto br-lan) confirm that ? (kernel 5.10). The ports acknowlege there is a link is up but no ping or other access for me.

olek210 · October 20, 2021, 8:04pm

Hi @wilsonyan,
I am the original author of these patches. This is an older patch version and it is broken. Here is a newer version of the patch. This patch probably doesn't improve performance. You will find more patches in my repository. Some are untested.

Combining 4 patches increases performance by 25%:

I will send patches upstream on the next net-next cycle.

jekkos · October 21, 2021, 7:05am

I also saw you have a patch to use 4 TX queues, was any testing done on that one? I'm seeing a frequent 'tx ring buffer full' error that sometimes stalls the ethernet connection. Any idea if this patch could help to resolve that? Thanks!

olek210 · October 21, 2021, 4:57pm

The patch for TX queue 4 has not been tested. 'Tx ring buffer full' messages appear on the console because DMA is slow. The system fills the descriptors faster than dma releases them. I think patches from this PR should solve this problem.

olek210 · October 21, 2021, 5:00pm

Using all the patches I was able to achieve 725 Mb / s upload and around 600 Mb / s download speed.

jekkos · October 21, 2021, 9:10pm

impressive. So in this case the 4 patches you mentioned in your earlier post? If I understand correctly then most of them were sent upstream and it might take some time until they are pulled in openwrt?

jekkos · October 22, 2021, 2:47pm

I have installed the patch you mentioned, and saw the message again today, this also gave me a short
hiccup in the network. Any other ideas on what I can try/test? maybe the IRQ balancing? I just saw your commit here which seems to contain all patches you previously mentioned.

Weird also is that it seemed to coincide with a dsl connection drop

[61678.440916] lantiq,xrx200-net 1e10b308.eth eth0: tx ring full

olek210 · October 23, 2021, 8:26pm

Are you connected to the router with a 100M link?

jekkos · October 23, 2021, 9:57pm

yes correct, it's an older PLC adapter to which the port is connected, only 100M.

olek210 · October 24, 2021, 6:30pm

None of my patches solve this problem. My patches increase the performance of the CPU port. Probably the faster the CPU port, the more frequently these messages appear on the console. This is because the CPU is connected to the switch with a 1GBps link, and your device is connected to the switch with a 100Mbps link. The CPU can send data faster than the switch forwards it over a slow link and the buffers run out quickly. There can be a maximum of 255 buffers per TX channel.

olek210 · November 16, 2021, 10:05pm

Good news. Some of the performance-enhancing patches were merged yesterday.

jekkos · November 19, 2021, 11:48am

@olek210 Does this branch contain all of these patches? I'd like to test them out and it would save me some time to apply the diffs to the openwrt repo myself.

olek210 · January 9, 2022, 12:05am

There is also a PR that increases efficiency. It contains the previously mentioned patches.

wilsonyan · January 9, 2022, 4:33am

im testing this currently, one thing I cant confirm is whether it's possible to have a port carry 2 tagged vlans on it

1 is ok, I haven't confirmed 2 work

olek210 · January 9, 2022, 12:39pm

Are you testing QinQ?

wilsonyan · January 9, 2022, 10:54pm

no, just basically want a public lan and a private lan with 1 port tagged in on both
maybe it's just the difficulty of DSA and UCI config or perhaps the vlan id was too high or something
i'll try again later

everything seems stable with your performance branch, been up for 3-4 days now no problems

olek210 · January 16, 2022, 9:37pm

My patches have been applied to the master in 7e484b9.

dietcoke73 · January 20, 2022, 9:28pm

Is that from WAN to LAN? I can only get 150mbps from wan to lan, on lan to lan I get 950 so there its great

olek210 · January 20, 2022, 9:47pm

Measurement results are between LAN and WAN. I can get 700Mbps with software flow offload enabled.