Re: [patch 02/10] x86/mce: Disable tracing and kprobes on do_machine_check()

[Peter Zijlstra]

On Wed, Feb 26, 2020 at 07:10:01AM -0800, Andy Lutomirski wrote:
> On Wed, Feb 26, 2020 at 5:28 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Tue, Feb 25, 2020 at 09:29:00PM -0800, Andy Lutomirski wrote:
> >
> > > >> +void notrace do_machine_check(struct pt_regs *regs, long error_code)
> > > >>  {
> > > >>    DECLARE_BITMAP(valid_banks, MAX_NR_BANKS);
> > > >>    DECLARE_BITMAP(toclear, MAX_NR_BANKS);
> > > >> @@ -1360,6 +1366,7 @@ void do_machine_check(struct pt_regs *re
> > > >>    ist_exit(regs);
> > > >>  }
> > > >>  EXPORT_SYMBOL_GPL(do_machine_check);
> > > >> +NOKPROBE_SYMBOL(do_machine_check);
> > > >
> > > > That won't protect all the function called by do_machine_check(), right?
> > > > There are lots of them.
> > > >
> > >
> > > It at least means we can survive to run actual C code in
> > > do_machine_check(), which lets us try to mitigate this issue further.
> > > PeterZ has patches for that, and maybe this series fixes it later on.
> > > (I'm reading in order!)
> >
> > Yeah, I don't cover that either. Making the kernel completely kprobe
> > safe is _lots_ more work I think.
> >
> > We really need some form of automation for this :/ The current situation
> > is completely nonsatisfactory.
> 
> I've looked at too many patches lately and lost track a bit of which
> is which.  Shouldn't a simple tracing_disable() or similar in
> do_machine_check() be sufficient?

It entirely depends on what the goal is :-/ On the one hand I see why
people might want function tracing / kprobes enabled, OTOH it's all
mighty frigging scary. Any tracing/probing/whatever on an MCE has the
potential to make a bad situation worse -- not unlike the same on #DF.

The same with that compiler instrumentation crap; allowing kprobes on
*SAN code has the potential to inject probes in arbitrary random code.
At the same time, if you're running a kernel with that on and injecting
kprobes in it, you're welcome to own the remaining pieces.

How far do we want to go? At some point I think we'll have to give
people rope, show then the knot and leave them be.

> We'd maybe want automation to check
> everything before it.  We still need to survive hitting a kprobe int3,
> but that shouldn't have recursion issues.

Right, so I think avoiding the obvious recursion issues is a more
tractable problem and yes some 'safe' spot annotation should be enough
to get automation working for that -- mostly.