just to make it clear... we have some theory that the nss firmware is setting up the switch PPE (packet processing engine)
But currently with this theory false, all the offload feature of the switch are NOT used.
In theory it should be possible to use them by using the ssdk api and the kernel offload api. But we have other priority and that is totally something that the community should handle by sideproject and would require lots of testing.
Btw I'm curious about what can those NPUs do? Seems that PPE could handle almost everything about networking.
they are the nss core... they pretty much do everything... they are driven by the nss firmware
Maybe this gives a more clear idea of PPE vs NPU (pretty big difference in Mp/s).
For 10GE, 14 Mp/s need to be processed, so the NPU cannot handle that by itself. Maybe it is used to set up what data goes to PPE.
Maybe someone else has a better idea.
Searching IPQ8078A bellow, there's a DS link to the device specification pdf, with detailed description of what NSS does:
https://wikidevi.wi-cat.ru/Qualcomm
Most probably on board there's "jumper" that may switch to load uboot from different address. But if you want to start play with it and be on safe side you may use TP-Link X60 GPL sources that contains uboot with full files including definition for ipq807x-hk02 board. The hardware is very similar to AX3600, clocks, memory and net interfaces are set. IMO. good starting point but if you don't have any fallback method you can very easy brick your device. Someone in the past already played with qca uboot: https://forum.archive.openwrt.org/viewtopic.php?id=65956&p=2 (see post 50) those flags looks promising but no guarantee that it will look exactly same in ipq807x... Generally without finding usb pin out (if even exists) or JTAG I wouldn't be so brave to test it
I've been using different robi releases since some months ago. All of them without problem, stable several weeks without any crash. But since two or three versions ago, the system is crashing the wifi each few days. I have several wifi "names" and all of them suddenly disappear.
Is not a memory problem, I've been monitoring memory and it's only using about a 60% more or less when the crash happens. A reboot of course fixes the issue.
Someone more is having this kind of problems with latest versions? I will try to get some logs, but is not easy, I don't have a laptop so I will need to move the big computer near the router the next time to try to get it.
Do you have any idea what version it started at?
Not too sure because it is not immediate, can pass two or three days before crashing, and the first time I ignored it thinking it was an sporadic crash.
It has happened three times in the latest 10 days more or less.
Consider that these build are from master so consider that some regression may also come from plain openwrt and not robi/ansuel repo
The recent airtime changes appear to have caused quite a few problems (crashes) for ath10k, they probably also caused havoc for ath11k. As these changes have been reverted in master a few hours ago, you could try that (pulling Robert's branch into current master, to get those commits).
There are also a lot of hostapd changes pulled to master a couple hours ago, large number of them are QCA specific: hostapd: update to 2022-06-02
We might need to wait for Robi to rebase against master, although the latest ipq-5-15-pr branch works fine for me, 6+ days uptime.
Hi Ansuel, I did some tests on the firmware https://github.com/Ansuel/openwrt/releases/tag/ipq807x-2022-05-22, it shows speed results better in conditions of interference or poor signal, access points were installed in the stadium in AC mode , at a frequency of 5745 MHz, the Xiomi AX 3600 worked at the same frequency, with external antennas. The devices connected in 802.11AX mode and got good speed and ping. Have you applied something in your firmware to work in conditions of interference or a weak signal ??? (please tell me what exactly), but there is a problem with memory growth.
Why did you stop releasing firmware builds?
got occupied by other project and also giving priority on searching bugs on plain ath11k without extra mod...
my repo have some qca patches and some tweak so that is probably it... mainly a reduction of some packet that probably increase perf in congested scenario (aka mgmt packet are smaller so they can be handled in a better way in your specific scenario)
It's a pity that you switched to another project, tell me it's possible to improve the work of Airtime to make it turn the queue faster and distribute the load between all clients, is this processed in the ath11k driver, or mac80211 ?. Tell me where you can change the "maximum number of connected devices" parameter?, maybe it's written in the sources ath11k driver?
And you tell me the meaning of life 
(that's a lot of question and the Airtime improvement stuff is still a nightmare i don't want to be part of... I already have qcom source)
As always these are community project so overloading devs with tons of question will end up exactly in the opposite of what someone would expect.
Ok, if not secret, where can I get the qcom source code from? . Can you share the code?
I count 4 seemingly QCA-specific commits of about 40. Not sure I'd consider that a large number but none of the commits looks particularly interesting anyway.
Those are just vendor specific commands, nothing else
I haven't looked at them too closely, but any of the generic part of that commit is not interesting at all? I have also seen Ansuel's commit on ath-master about a leak fixed that is not yet part of your tree:
ath11k: fix missing skb drop on htc_tx_completion error
