Re: HEADS UP! Do not update on main currently (panic - on boot)

On Thu, 6 Mar 2025, Baptiste Daroussin wrote:

> On Wed 05 Mar 23:39, Bjoern A. Zeeb wrote:
>> Hi,
>>
>> the fix from https://cgit.FreeBSD.org/src/commit/?id=27bf5c405bf2eb69392e45c06605defc78882612
>> makes wireless panic for me in my development branch;  that said I am
>> working on TKIP currently so I have code in progress.
>>
>> In case you hit the panic (on boot) because you have updated, please
>> disable the HW_CRYPTO tunable from loader and live without HT/VHT for a
>> few hours again.
>>
>> I am looking where the problem comes from now and for a fix then.
>>
>> /bz
>>
>
> Thanks for the heads up!
>
> The weekly snapshots for pkgbase are generated the weekend, if you haven't been
> able to figure out a fix by then, if would be nice for pkgbase consumers if you
> could revert the change before the next snapshot building.
>
> Many people took the habbit to run pkgbase from weekly snapshots and reboot
> every monday on a fresh new current machine.

I have a very crude workaround I am testing.  I still hope to be able to
fix it more properly if net80211 locking allows us.

The problem isn't so much of reverting.  11db70b6057e41b259dc2245cd893d5b19179fcc
taked about the possible panic on key delete I could not reproduce
anymore.  As I had guessed the mentioned commit was the reason.  The
problem was that that commit had logic inverted and we corrected that
yesterday so now net80211 and ath AP mode and others are fine again but
the panic on key delete is back depending on timing.

The problem is that net80211 already has incosistent locking there and
we have to conditionally unlock for the downcall to driver/firmware
which can sleep.
Both together opens a possible race which is there in theory with or
without LinuxKPI as it is two threads (wpa and net80211) racing.
I am just able to trigger it again (note that some drivers simply omitted
these bits; that's why you don't see ic_cryptocaps set in iwm or iwn)
but that no longer works as firmware needs to be able to handle crypto
for doing A-MPDU offloading.

HTH to understand,
/bz

-- 
Bjoern A. Zeeb                                                     r15:7