Re: perf/tracepoint: another fuzzer generated lockup

[Vince Weaver <vincent.weaver@maine.edu>]

On Thu, 14 Nov 2013, Peter Zijlstra wrote:

> On Thu, Nov 14, 2013 at 12:20:35PM -0500, Vince Weaver wrote:
> > 
> > possibly not an issue with your patch, but found in nearby code:
> > 
> > in perf_trace_init:
> > 
> >         int event_id = p_event->attr.config;
> > 
> > 	...
> > 
> >                 if (tp_event->event.type == event_id &&
> > 
> > 
> > note that "attr.config" is 64-bit but "event_id" is only 32-bit,
> > so things like 0xffffffff00000018 will map to irq_work_exit.
>
> I will have a look.

Assuming this is a real issue, I'll submit a patch to fix this if I'm not
beaten to it by someone else.

> > I'm not sure if your PERM code properly handles this case, I don't have 
> > time to work through the macro-magic in your patch.
> 
> Can't blame you; took me a fair chunk of the day to come up with this.
> That trace event stuff is horrid.

I did some tests both with the high bits set and non-set and wasn't able 
to trigger the bug.  Also ran the fuzzer for a while and while it keeps 
finding other bugs, it hasn't tripped over any that were overtly 
tracepoint related. So you can add my

Tested-by: Vince Weaver <vincent.weaver@maine.edu>

Vince