arm64: dts: rockchip: Change ethernet TXD timing delay value
Subsystems (2)
PATCH v2 1/1 arm64: dts: rockchip: Change ethernet TXD timing delay value
Very poor network performance is observed on the ODROID-M1 Board (REV 1.0 20220613) when running iperf3 (under 200Mbits/sec), and this problem can be resolved after changing the value of TXD timing to 0x35 from 0x4f.
--- arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts +1 -1
@@ -210,7 +210,7 @@ &gmac0_rgmii_clk &gmac0_rgmii_bus>; status = "okay"; - tx_delay = <0x4f>; + tx_delay = <0x35>; rx_delay = <0x2d>; };
Signature · 1 line
2.34.1
Hi,
Quoted 4 lines
Very poor network performance is observed on the ODROID-M1 Board (REV 1.0 20220613) when running iperf3 (under 200Mbits/sec), and this problem can be resolved after changing the value of TXD timing to
0x35from0x4f.
Can you try changing phy-mode to "rgmii-id" ?
```
--- a/arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts
+++ b/arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts
@@ -200,7 +200,7 @@ &gmac0 {
assigned-clock-rates = <0>, <125000000>;
clock_in_out = "output";
phy-handle = <&rgmii_phy0>;
- phy-mode = "rgmii";
+ phy-mode = "rgmii-id";
phy-supply = <&vcc3v3_sys>;
pinctrl-names = "default";
pinctrl-0 = <&gmac0_miim
@@ -209,9 +209,6 @@ &gmac0_rx_bus2
&gmac0_rgmii_clk
&gmac0_rgmii_bus>;
status = "okay";
-
- tx_delay = <0x4f>;
- rx_delay = <0x2d>;
};
&gpu {
```--
2.25.1
Quoted 33 lines · Chukun Pan
Hi,
Quoted 4 lines
Very poor network performance is observed on the ODROID-M1 Board (REV 1.0 20220613) when running iperf3 (under 200Mbits/sec), and this problem can be resolved after changing the value of TXD timing to
0x35from0x4f.Can you try changing phy-mode to "rgmii-id" ?
```
arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts +1
-4@@ -200,7 +200,7 @@ &gmac0 { assigned-clock-rates = <0>, <125000000>; clock_in_out = "output"; phy-handle = <&rgmii_phy0>; - phy-mode = "rgmii"; + phy-mode = "rgmii-id"; phy-supply = <&vcc3v3_sys>; pinctrl-names = "default"; pinctrl-0 = <&gmac0_miim@@ -209,9 +209,6 @@ &gmac0_rx_bus2 &gmac0_rgmii_clk &gmac0_rgmii_bus>; status = "okay"; - - tx_delay = <0x4f>; - rx_delay = <0x2d>; }; &gpu {```
Quoted 4 lines
--
2.25.1
Thank you for suggesting me to try 'rgmii-id' again.
I'd remember that 'rgmii-id' didn't work with the very early version of the board, however your suggestion works. Let me test the change with more boards and submit the new patch when it's done.
Thanks a lot.
Dongjin Kim.
Miguel Jesus Seron Gonzalez
Tested-byHi Dongjin,
Quoted 5 lines · Dongjin Kim
Thank you for suggesting me to try 'rgmii-id' again.
I'd remember that 'rgmii-id' didn't work with the very early version of the board, however your suggestion works. Let me test the change with more boards and submit the new patch when it's done.
I ran into the same problem on an ODROID-M1 board marked "REV 1.0 20211112", an earlier revision than the 20220613 one in your patch, running Home Assistant OS (kernel 6.18.52). With the current tx_delay = <0x4f>, about 6.6% of all TCP segments were retransmitted at 1000 Mbit/s and larger HTTP transfers stalled.
I tested Chukun's suggestion (phy-mode = "rgmii-id", with tx_delay and rx_delay removed; the driver reports "init for RGMII_ID") on the same board, on a 1000 Mbit/s full-duplex link. It works fine on this early revision as well:
- 3 retransmitted TCP segments out of 24929 (0.012%) over 3 h 49 min of uptime, and no RX/TX errors
- 90 sequential requests for large files (22.5 MB in total): no failures, 1.4 s in total, the slowest one 47 ms
- 120 parallel requests (75 MB in total): no failures, 1.2 s
For comparison, tx_delay = <0x35> gave 0.03% retransmissions on the same board.
Are you still planning to send a v3 with rgmii-id? If you don't have time for it, I'd be happy to send it with Chukun's Suggested-by. Either way, for the rgmii-id version:
Thanks,
Miguel