Re: [RFC PATCH] dynamic debug - adding ring buffer storage support

[Jason Baron <jbaron@redhat.com>]

On Wed, Dec 30, 2009 at 08:44:17PM -0500, Steven Rostedt wrote:
> 
> > That said, I sometimes dream about one event per printk.
> > 
> > Having, say:
> > 
> > /debug/tracing/events/printk/
> >          |
> >          ---- kernel/
> >          |      |
> >          |      ------- time/
> >          |      |        |
> >          |      |        ---- clocksource.c
> >          |      |                 |
> >          |      |                 --- clocksource_unstable:218/
> >          |      |                 |            |
> >          |      |                 |            ---- format
> >          |      |                 |            |
> >          |      |                 |            ---- filter
> >          |      |                 |            |
> >          |      |                 |            ---- enable
> >          |      |                 --- [...]
> >          |      ------- [...]
> >          |
> >          ---- drivers/
> >          |       |
> >          |       ---- [...]
> >          |
> >          ---- [...]
> > 
> > 
> > That would give a total control over every printk, trace_printk, etc...
> > 
> > Too bad that would bloat the memory.
> > Well, that could be wrapped in a single, wildly implemented (understand:
> > not using TRACE_EVENT macro) trace event, something able to walk through
> > every calls of printk, trace_printk, early_printk, etc... and imitate
> > a per printk event granularity.
> > 
> > But still it needs to be useful...
> 
> 
> I think we can do the above without bloating memory. Yes we would not
> need the TRACE_EVENT macro for this. The TRACE_EVENT macro is just for
> generic tracing, but we could easily come up with something specific for
> the printk's that will not bloat the kernel as much.
> 
> When I get some time, I may try to play with this idea.
> 
> -- Steve
> 

I agree with this direction...in terms of the implementation I was
thinking it could be very similar to the tracepoint optimization i've
been working on. Where we basically end up with just a 'nop' in place of
the printk and then when we enable it we patch the 'nop' with a jump to
the proper printk location...

-Jason