Re: Dual-licensing LTTng, marker and tracepoints under GPLv2+/LGPLv2.1+

[Mark Wielaard <mjw@redhat.com>]

Mathieu Desnoyers <compudj <at> krystal.dyndns.org> writes:
> * Frank Ch. Eigler (fche <at> redhat.com) wrote:
> > > [...]The goal is to permit this library, which includes the
> > > tracepoints, markers and LTTng features, to be used by userspace
> > > applications and libraries so they can add static instrumentation
> > > (as we currently do in the kernel). [...]
> > 
> > Just in case you're not aware, some API prior art for this is the
> > dtrace sdt.h widget (already supported by systemtap), and of course
> > it has no similarity to the various kernel tracing APIs.
> 
> Yep, we're aware of this. However, last time I checked, dtrace SDT used
> a breakpoint even for their userspace tracing, which has a way too large
> performance overhead for our needs.

You might want to check again. Although the systemtap sdt.h should be source
compatible with dtrace markers since it provides the same compile time
interfaces/macros (which is how systemtap support java hotspot probing for
example), it isn't tied to any implementation details. The default
implementation inserts a NOP that can be replaced by a breakpoint instruction.
But Stan Cox added variants that can use syscalls/utrace/kprobes which have
different performance overhead tradeoffs
(http://article.gmane.org/gmane.linux.systemtap/12667). This is an
implementation detail though since the source interfaces are the same. Just
recompiling your application and/or libraries against a newer version of the
sdt.h interfaces will automagically provide the different probe types.

Cheers,

Mark