Re: question on system tap

[fche at redhat dot com (Frank Ch. Eigler)]

"Hebenstreit, Michael" <michael.hebenstreit@intel.com> writes:

> [...]
> a) I start with instrumenting a kernel function A
> b) I run my benchmark
> c) the output should tell me which functions were CALLED by A and the timestamp when they return to A
> d) for those function that interest me more I repeat this procedure
> this way my instrumentation is kept light; I could not find any way to do this with systemtap (maybe a different tool would be better?)

Function-by-function tracing is by nature expensive, even with tools
other than systemtap.  Something like para-callgraph.stp should give
some useful data.

http://sourceware.org/systemtap/examples/keyword-index.html#CALLGRAPH

You'd identify function A with the second parameter, and all the
candidate functions in the call graph with the first parameter, as in:

# stap para-callgraph.stp 'kernel.function("*@fs/*")' 'module("lustre").function("A")'

If you want to designate the candidate functions with more than one
probe point, as in kernel.something as well as
module/lustre.something, then you'll need to modify the script and
basically hand-substitute in $1.  Same for the enable/disable trigger
$2.

This report gives you a lot of information, including parameter
values.  If you want less info (thus let the script run faster and
smaller), you can simplify the printf() in the script to remove
$$parms.

See also the older version of this script with a bit more narrative:
http://sourceware.org/systemtap/wiki/WSCallGraph

- FChE