Re: selftests: net: udpgso_bench.sh: RIP: 0010:lookup_reuseport

[Arnd Bergmann]

On Sat, May 27, 2023, at 05:49, Kuniyuki Iwashima wrote:
> From: Jakub Kicinski <kuba@xxxxxxxxxx>
> Date: Fri, 26 May 2023 20:16:07 -0700
>> On Wed, 24 May 2023 13:24:15 +0530 Naresh Kamboju wrote:
>> > While running selftests: net: udpgso_bench.sh on qemu-x86_64 the following
>> > kernel crash noticed on stable rc 6.3.4-rc2 kernel.
>> 
>> Can you repro this or it's just a one-off?
>> 
>> Adding some experts to CC.
>
> FWIW, I couldn't reproduce it on my x86_64 QEMU setup & 6.4.0-rc3
> at least 5 times, so maybe one-off ?

This looks like one of several spurious reports that lkft has produced
recently, where an 'int3' trap instruction is executed in a function
that is live-patched, but at a point where the int3 is not expected.

Anders managed to get a reproducer for one of these on his manchine
yesterday, and has narrowed it down to failing on qemu-7.2.2 but
not failing on qemu-8.0.

The current theory right now is that this is a qemu bug when
dealing with self-modifying x86 code that has been fixed in
qemu-8.0 already, and my suggestion would be to ignore all bugs
found by lkft that involve an 'int3' trap, and instead change
the lkft setup to use either qemu-8.0 or run the test systems
in kvm (which would also be much faster and save resources).

Someone still needs to get to the bottom of this bug to see
if it's in qemu or in the kernel livepatching code, but I'm
sure it has nothing to do with the ipv6 stack.

      Arnd