cvs commit: src/sys/net if_vlan.c

[Yar Tikhiy]

On Fri, Aug 04, 2006 at 04:44:07PM -0400, John Baldwin wrote:
> 
> To be honest, as someone who works with bug reports, I'd actually like 
> backtraces up front w/o requiring the user to compile a custom kernel, etc.  
> Having a simple backend in place and kdb_backtrace()'s where relevant would 
> be very handy. :)
> 
> > > Places that call kdb_enter() aren't all #ifdef KDB IIRC.  It's 
> > > just a feature that kdb_foo() functions become NOPs when the kernel isn't 
> > > configured for debugging, so I think the #ifdef KDB's would be redundant.
> > 
> > None of the kdb_*() functions in src/sys/kern/subr_kdb.c turn into
> > NOPs when option KDB is not present. They are all unconditionally
> > functional by design and should therefore be called conditionally
> > by consequence.
> 
> Well, given that separation, I'm not sure KDB is the right option to make 
> calls conditional.  Rather, some specific is-debugging-enabled? option (like 
> INARIANTS or FOO_DEBUG) should be used instead.  i.e.:
> 
> #ifdef FOO_DEBUG
> 	if (foo_bad) {
> 		printf("foo is bad\n");
> 		kdb_backtrace();
> 	}
> #endif
> 
> I don't think that warrants an extra #ifdef KDB.

Please excuse me, but there is a small inconsistency in your words.
On the one hand, you wish users could obtain and post backtraces
with no special efforts.  This is a great point because users don't
always have time or resources to reproduce a problem with kernel
debug features enabled, and some weird problems defy reproducing.
On the other hand, you suggest putting kdb_backtrace() calls under
secial #ifdef's.  That would effectively cancel out the benefits
from using kdb_backtrace() for "mild debugging" because you would
still have to have the users re-compile their kernels or modules
and try to catch the bug again.  A call to kdb_backtrace() is cheap,
so there is little sense in leaving it out from production kernels
and modules.  IMHO the only case when it should be done is when the
consistency check around kdb_backtrace() is expensive and sits on
a performance-critical path.

-- 
Yar