I would bet those NSS-clients are required for VLAN and bridge.
VLAN one is for sure, bridge one on the other hands does HW offloading to the PPE.
That is even better.
Can you make them on the regular restart branch?
That one pulls the NSS packages from the feed, other one is archive only.
@Ansuel Reverting the change in core patch makes no difference.
I am kind of stupid, CPU load is actually rather low, its just iperf generating traffic at 5-6% on all cores.
So it would mean its offloading.
But weirdly it has TCP retries while running of an external device has 0 retries..
It would be beneficial, as it would offload bridging to the HW.
Honestly, switchdev was made for this kind of stuff where you can offload various stuff to the HW.
But I wouldn't call it must have.
The more I think about it feels like the time could be better spent on cleanup and upstreaming.
[ 0.179380] qcom-qmp-phy 84000.phy: Registered Qcom-QMP phy
[ 0.181137] qcom-pcie 20000000.pci: host bridge /soc/pci@20000000 ranges:
[ 0.181188] qcom-pcie 20000000.pci: IO 0x0020200000..0x00202fffff -> 0x0020200000
[ 0.181215] qcom-pcie 20000000.pci: MEM 0x0020220000..0x002fffffff -> 0x0020220000
[ 1.300917] qcom-pcie 20000000.pci: Phy link never came up
[ 1.301074] qcom-pcie 20000000.pci: PCI host bridge to bus 0000:00
[ 1.301091] pci_bus 0000:00: root bus resource [bus 00-ff]
[ 1.301105] pci_bus 0000:00: root bus resource [io 0x0000-0xfffff] (bus address [0x20200000-0x202fffff])
[ 1.301116] pci_bus 0000:00: root bus resource [mem 0x20220000-0x2fffffff]
[ 1.301161] pci 0000:00:00.0: [17cb:1002] type 01 class 0x060400
[ 1.301185] pci 0000:00:00.0: reg 0x10: [mem 0x00000000-0x00000fff]
[ 1.301265] pci 0000:00:00.0: PME# supported from D0 D3hot D3cold
[ 1.306690] pci 0000:00:00.0: BAR 8: assigned [mem 0x20300000-0x204fffff]
[ 1.306710] pci 0000:00:00.0: BAR 9: assigned [mem 0x20500000-0x206fffff 64bit pref]
[ 1.306723] pci 0000:00:00.0: BAR 0: assigned [mem 0x20220000-0x20220fff]
[ 1.306736] pci 0000:00:00.0: BAR 7: assigned [io 0x1000-0x1fff]
[ 1.306749] pci 0000:00:00.0: PCI bridge to [bus 01-ff]
[ 1.306760] pci 0000:00:00.0: bridge window [io 0x1000-0x1fff]
[ 1.306772] pci 0000:00:00.0: bridge window [mem 0x20300000-0x204fffff]
[ 1.306784] pci 0000:00:00.0: bridge window [mem 0x20500000-0x206fffff 64bit pref]
[ 1.306808] pci 0000:00:00.0: Max Payload Size set to 128/ 128 (was 128), Max Read Rq 128
[ 1.308676] pcieport 0000:00:00.0: PME: Signaling with IRQ 103
[ 1.309761] bam-dma-engine 704000.dma: num-channels unspecified in dt
[ 1.309778] bam-dma-engine 704000.dma: num-ees unspecified in dt
The wifi radio is not visible though...
Is there really no easy way to recompile the dts and make an initramfs from it only without deleting and recompiling the entire kernel every time? That is such a waste of time
Also one of my ethernet ports is causing a warning, connecting a cable causes:
Nope, you can see that phy link never came up.
You can use lspci to see what the kernel sees.
I am gonna guess that the radio is on the Gen3 port, that is &pcie0.
perst-gpio for it is: perst-gpio = <&tlmm 58 GPIO_ACTIVE_LOW>;
If that does not cause a PHY linkup then you can try &pcie1
perst-gpio for it is: perst-gpio = <&tlmm 61 GPIO_ACTIVE_LOW>;
But for it, you also must enable the &qmp_pcie_phy1
Then simply use lspci to see what pops up.
Radio is not gonna work just by itself, you gotta enable the pci ath11k kmod as well and the firmware for QCA6390 and hope that your QCA6290 is close enough and supported.
Sure looks like it, 16MB of memory used more.
Can you try doing a crazy thing and connecting a client to WLAN, running speedtest and then checking memory usage?
this is before the speedtest so not related... currently i have my phone and my laptop connected to the router (wifi with no additional patch so no offload and all)
But ath11k calls ioremap like only 2 times.
On start it uses the devmem ioremap for the base resource from DTS, and once more for the BDF.
BTW, I am doing 443/92 Mbits on HE40 5G radio on my phone.
Offloading NAT really helped as now CPU load will be like 60% on one core and 15% on others.
Note, that this is not iperf3, but speedtest to my ISP IXP, so its topping out my uplink at this point as the network gets really loaded and I cant get the 500/120Mbits like in other times of the day.
iperf from the phone to desktop is still horrendus at 40Mbits.
@Ansuel My RAM consumption was 134MB, and then after starting ath11k and connecting a client it went to 145MB.
And now its just climbing at 200+MB in like 2 minutes.
So it looks like its leaking badly, like you can see it with your eyes.