Re: BUG: unable to handle kernel paging request in bpf_dispatcher_xdp

[Jiri Olsa]

On Fri, Dec 09, 2022 at 12:22:37PM +0100, Jiri Olsa wrote:

SBIP

> > > > > > >
> > > > > > > I'm trying to understand the severity of the issues and
> > > > > > > whether we need to revert that commit asap since the merge window
> > > > > > > is about to start.
> > > > > > 
> > > > > > Jiri, Peter,
> > > > > > 
> > > > > > ping.
> > > > > > 
> > > > > > cc-ing Thorsten, since he's tracking it now.
> > > > > > 
> > > > > > The config has CONFIG_X86_KERNEL_IBT=y.
> > > > > > Is it related?
> > > > > 
> > > > > sorry for late reply.. I still did not find the reason,
> > > > > but I did not try with IBT yet, will test now
> > > > 
> > > > no difference with IBT enabled, can't reproduce the issue
> > > > 
> > > 
> > > ok, scratch that.. the reproducer got stuck on wifi init :-\
> > > 
> > > after I fix that I can now reproduce on my local config with
> > > IBT enabled or disabled.. it's something else
> > 
> > I'm getting the error also when reverting the static call change,
> > looking for good commit, bisecting
> > 
> > I'm getting fail with:
> >    f0c4d9fc9cc9 (tag: v6.1-rc4) Linux 6.1-rc4
> > 
> > v6.1-rc1 is ok
> 
> so far I narrowed it down between rc1 and rc3.. bisect got me nowhere so far
> 
> attaching some more logs

looking at the code.. how do we ensure that code running through
bpf_prog_run_xdp will not get dispatcher image changed while
it's being exetuted

we use 'the other half' of the image when we add/remove programs,
but could bpf_dispatcher_update race with bpf_prog_run_xdp like:


cpu 0:                                  cpu 1:

bpf_prog_run_xdp
   ...
   bpf_dispatcher_xdp_func
      start exec image at offset 0x0

                                        bpf_dispatcher_update
                                                update image at offset 0x800
                                        bpf_dispatcher_update
                                                update image at offset 0x0

      still in image at offset 0x0


that might explain why I wasn't able to trigger that on
bare metal just in qemu

jirka