Re: Dtrace panic'ed

At Sat, 14 Feb 2009 08:25:02 -0500,
Andrew Gallatin wrote:
> 
> Robert Watson wrote:
> > 
> > On Fri, 13 Feb 2009, Andrew Gallatin wrote:
> > 
> >> I was trying to run a simple dtrace profiling script, and panic'ed the 
> >> machine using today's -current on an 8-way opteron.
> > 
> > Oh, you actually got a panic using the profile provider?  For me the box 
> > appears to go into la-la land.  I've also seen a few double faults using 
> > fbt a bit too gratuitously.  Unfortunately, I haven't found time to 
> > debug it, but I wondered if perhaps DTrace is not recognizing its own 
> > functions (i.e., ones it shouldn't try to trace) properly and/or failing 
> > to disable interrupts or enter a critical section at important moments.
> 
> I think that sounds like a likely hypothesis.  It is a shame it doesn't
> work.  Dtrace's profile provider is so useful...
> 
> BTW, did you see my next message where I was trying to
> find the null pointer?  It seemed almost as if the state saved
> by the kernel crashdump was "correct" (eg, no NULL pointer),
> but the code was executing with a different, corrupt register set.
> It may just be my misunderstanding on amd64 asm though.
> 

For me I get the same result as Robert.  In particular it is ONLY with
the profile provider but not the other timing provider (tick).  And it
is also, for me, only near the 100 Hz profile.

> As another aside, what is up with kgdb & module debug these days?
> It seems to load module symbols automagically these days,
> which is very cool.  But the modules themselves do not seem
> to have symbols that gdb understands, even though I built
> with 'makeoptions     DEBUG=-g'

This is a side effect of the ctfconvert.  John Baldwin and I fixed
this for the kernel, but I have not yet fixed it for modules.

Best,
George