Re: linux-next: stall warnings and deadlock on Arm64 (was: [PATCH] kfence: Avoid stalling...)

[Mark Rutland <mark.rutland@arm.com>]

On Fri, Nov 20, 2020 at 09:38:24AM -0800, Paul E. McKenney wrote:
> On Fri, Nov 20, 2020 at 03:22:00PM +0000, Mark Rutland wrote:
> > On Fri, Nov 20, 2020 at 06:39:28AM -0800, Paul E. McKenney wrote:
> > > On Fri, Nov 20, 2020 at 03:19:28PM +0100, Marco Elver wrote:
> > > > I found that disabling ftrace for some of kernel/rcu (see below) solved
> > > > the stalls (and any mention of deadlocks as a side-effect I assume),
> > > > resulting in successful boot.
> > > > 
> > > > Does that provide any additional clues? I tried to narrow it down to 1-2
> > > > files, but that doesn't seem to work.
> > > 
> > > There were similar issues during the x86/entry work.  Are the ARM guys
> > > doing arm64/entry work now?
> > 
> > I'm currently looking at it. I had been trying to shift things to C for
> > a while, and right now I'm trying to fix the lockdep state tracking,
> > which is requiring untangling lockdep/rcu/tracing.
> > 
> > The main issue I see remaining atm is that we don't save/restore the
> > lockdep state over exceptions taken from kernel to kernel. That could
> > result in lockdep thinking IRQs are disabled when they're actually
> > enabled (because code in the nested context might do a save/restore
> > while IRQs are disabled, then return to a context where IRQs are
> > enabled), but AFAICT shouldn't result in the inverse in most cases since
> > the non-NMI handlers all call lockdep_hardirqs_disabled().
> > 
> > I'm at a loss to explaim the rcu vs ftrace bits, so if you have any
> > pointers to the issuies ween with the x86 rework that'd be quite handy.
> 
> There were several over a number of months.  I especially recall issues
> with the direct-from-idle execution of smp_call_function*() handlers,
> and also with some of the special cases in the entry code, for example,
> reentering the kernel from the kernel.  This latter could cause RCU to
> not be watching when it should have been or vice versa.

Ah; those are precisely the cases I'm currently fixing, so if we're
lucky this is an indirect result of one of those rather than a novel
source of pain...

> I would of course be most aware of the issues that impinged on RCU
> and that were located by rcutorture.  This is actually not hard to run,
> especially if the ARM bits in the scripting have managed to avoid bitrot.
> The "modprobe rcutorture" approach has fewer dependencies.  Either way:
> https://paulmck.livejournal.com/57769.html and later posts.

That is a very good idea. I'd been relying on Syzkaller to tickle the
issue, but the torture infrastructure is a much better fit for this
problem. I hadn't realise how comprehensive the scripting was, thanks
for this!

I'll see about giving that a go once I have the irq-from-idle cases
sorted, as those are very obviously broken if you hack
trace_hardirqs_{on,off}() to check that RCU is watching.

Thanks,
Mark.