Re: [PATCH][RFC] Infrastructure for compact call location representation

[Nick Piggin <npiggin@suse.de>]

On Wed, Jun 09, 2010 at 11:22:44AM -0700, David VomLehn wrote:
> On Wed, Jun 09, 2010 at 03:44:17AM -0500, Nick Piggin wrote:
> > On Tue, Jun 08, 2010 at 08:44:56AM -0700, Stephen Hemminger wrote:
> > > On Mon, 7 Jun 2010 17:30:52 -0700
> > > David VomLehn <dvomlehn@cisco.com> wrote:
> > > > History
> > > > v2	Support small callsite IDs and split out out-of-band parameter
> > > > 	parsing.
> > > > V1	Initial release
> > > > 
> > > > Signed-off-by: David VomLehn <dvomlehn@cisco.com>
> > > 
> > > This is really Linux Kernel Mailing List material (not just netdev). And it will
> > > be a hard sell to get it accepted, because it is basically an alternative call
> > > tracing mechanism, and there are already several of these in use or under development
> > > (see perf and ftrace).
> > 
> > What about a generic extension or layer on top of stacktrace that
> > does caching and unique IDs for stack traces. This way you can get
> > callsites or _full_ stack traces if required, and it shouldn't require
> > any extra magic in the net functions.
> 
> Since the code calls BUG() when it detects an error, you already get the
> full stack trace of the location where the problem is detected. The question
> is the relative cost and benefits of a full stack trace of the previous
> sk_buff state modification. Since I'm working in a MIPS processor
> environment, I am rather prejudiced against doing any stack trace I don't
> have to; for now, at least, they are *very* expensive on MIPS.

Point is that you could select this depending on whether or not you
want it. If not, then you can just record the current IP.

 
> The two times this code (or its ancestor) has found problems in a deployed
> software stack, the engineers reported they there were able to immediately
> find and fix the problem. This suggests that we don't need to take on the
> complexity of the stack backtrace, at least for now. If this gets added to
> the mainline and people find they need the extra information, I'd be all
> for it.

I don't think it would get added to mainline with the tracing stuff
as a special hack under net/. I'm not saying it's not useful, but it
should just go into core code.