Hi,

Some instrumentation tools on Linux, like Itrace and systemtap
(http://sourceware.org/systemtap) now use the kprobe infrastructure to
gather information. One of the requirements of projects like systemtap 
is the ability to define multiple kprobes at a given address.

To this end, here is a patch that provides the feature. Patch is
against linux-2.6.12-rc2.

This patch provides the facility to register multiple kprobes at the
same address using the existing interfaces. The house-keeping in
case of multiple kprobes is taken care of within the base kprobes
infrastructure.

Another approach considered was to have a layer above the existing
kprobes base (no modification to current kprobes base at all). A patch
to this end is available at:

http://sourceware.org/ml/systemtap/2005-q2/msg00089.html

This approach would also allow for two sets of interfaces for
un/registering kprobes. There has been quite a few discussions on the
systemtap lists whether two interfaces are necessary or not. But, with
the current kprobes locking model, the layered approach leaves room
for a few kprobe registration races.

Both approaches are architecture agnostic.

Other kprobe enhancements in the pipeline, such as, improving kprobe
locking, support for return address probes, etc. Given that, the main
questions to be answered now for the multi-kprobe feature are:

1. Is the approach taken by the patch attached good?
2. Do we take the layered approach?
3. If we take the layered approach, do we need multiple interfaces?
4. Would existing users be impacted by the change?

Comments?

Thanks,
Ananth