On Mon, Nov 23, 2020 at 07:32PM +0000, Mark Rutland wrote:
> On Fri, Nov 20, 2020 at 03:03:32PM +0100, Marco Elver wrote:
> > On Fri, Nov 20, 2020 at 10:30AM +0000, Mark Rutland wrote:
> > > On Thu, Nov 19, 2020 at 10:53:53PM +0000, Will Deacon wrote:
> > > > FWIW, arm64 is known broken wrt lockdep and irq tracing atm. Mark has been
> > > > looking at that and I think he is close to having something workable.
> > > > 
> > > > Mark -- is there anything Marco and Paul can try out?
> > > 
> > > I initially traced some issues back to commit:
> > > 
> > >   044d0d6de9f50192 ("lockdep: Only trace IRQ edges")
> > > 
> > > ... and that change of semantic could cause us to miss edges in some
> > > cases, but IIUC mostly where we haven't done the right thing in
> > > exception entry/return.
> > > 
> > > I don't think my patches address this case yet, but my WIP (currently
> > > just fixing user<->kernel transitions) is at:
> > > 
> > > https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/irq-fixes
> > > 
> > > I'm looking into the kernel<->kernel transitions now, and I know that we
> > > mess up RCU management for a small window around arch_cpu_idle, but it's
> > > not immediately clear to me if either of those cases could cause this
> > > report.
> > 
> > Thank you -- I tried your irq-fixes, however that didn't seem to fix the
> > problem (still get warnings and then a panic). :-/
> 
> I've just updated that branch with a new version which I hope covers
> kernel<->kernel transitions too. If you get a chance, would you mind
> giving that a spin?
> 
> The HEAD commit should be:
> 
>   a51334f033f8ee88 ("HACK: check IRQ tracing has RCU watching")

Thank you! Your series appears to work and fixes the stalls and
deadlocks (3 trials)! I noticed there are a bunch of warnings in the log
that might be relevant (see attached).

Note, I also reverted

  sched/core: Allow try_invoke_on_locked_down_task() with irqs disabled

and that still works.

Thanks,
-- Marco

> Otherwise, I intend to clean that up and post it tomorrow (without the
> additional debug hacks). I've thrown my local Syzkaller instance at it
> in the mean time (and if I get the chance tomrrow I'll try to get
> rcutorture setup), and the only report I'm seeing so far looks genuine:
> 
> | BUG: sleeping function called from invalid context in sta_info_move_state
> 
> ... as that was reported on x86 too, per:
> 
> https://syzkaller.appspot.com/bug?id=6c7899acf008be2ddcddb46a2567c2153193632a
> 
> Thanks,
> Mark.