So after a few days of poking at it, I .. kinda think I know what's going on with wifi association on this realtek USB driver in FreeBSD. It /looks/ like the NIC isn't fully ready to send frames when the association frame is sent. I don't know exactly what "ready" entails though. - ThreadSky

erikarn.bsky.social • 40 days ago

So after a few days of poking at it, I .. kinda think I know what's going on with wifi association on this realtek USB driver in FreeBSD.

It /looks/ like the NIC isn't fully ready to send frames when the association frame is sent. I don't know exactly what "ready" entails though.

Comments

snowman.bsky.social•39 days ago

TMI

erikarn.bsky.social•40 days ago

A 10ms sleep after programming the channel info is enough. But why 10ms? Why not 5ms? I'm not sure. Why's it only happen on newer laptops on USB-3, and not older ones with USB-3? There's more CPU cores available and they're faster - so stuff can race between threads/CPUs.

sneakyness.com•40 days ago

Only thing I could think of is maybe the newer ones are 3.1+ and are doing some sort of additional discovery/negotiation for the 240w power delivery spec?

erikarn.bsky.social•40 days ago

The NIC is USB-2, so it's not doing extra negotiation!

I looked at the timing a while back and the newer laptops are able to do the control transfers (used for register IO) at the spec minimum (125us). The older laptops can't. So, I thought it was that, and we needed more sleeps after each write.

erikarn.bsky.social•40 days ago

But then I tried to just add sleeps manually between steps when programming the channel, not in the register IO path itself. It turns out it needs a long sleep (10ms) at the end, or "enough" sleeps of 1ms each in between steps.

So yeah, I think it's something going on inside the FW in the NIC.

erikarn.bsky.social•40 days ago

Like, the RTL8192CU and RTL8812AU/RTL8821AU doesn't need the sleep. The RTL8192EU / RTL8188EU does. The firmware may be doing something to do with power saving, some calibration may be going on, something to do with the TXPAUSE register (which pauses TX queues), etc.

I need much more tracing!

sneakyness.com•40 days ago

This sounds maddening, the only other straw I can think of to grasp at is autosuspend power saving somehow fucking with it? I think the default is 2 seconds

good luck dude

bonniezilla.bsky.social•40 days ago

NERD :)

erikarn.bsky.social•40 days ago

.. and to make things worse, Linux + mac80211 + wpa_supplicant resends the auth frame 3 times itself. FreeBSD's net80211 doesn't do that.

So, I think the 10ms sleep will need to go in, with a big ass comment explaining stuff, and then I'll have to revisit things once some more clean-ups are done.

erikarn.bsky.social•40 days ago

(and if you're curious of the shenanigans going on - https://reviews.freebsd.org/D48517 and https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=247528#c60. Fun times.)

Comments

Posting Rules

Reply