Re: KPROBES: Instrumenting a function's call site

[Ananth N Mavinakayanahalli <ananth@in.ibm.com>]

On Tue, Sep 25, 2007 at 06:12:38PM -0400, Avishay Traeger wrote:
> Hello,
> I am trying to use kprobes to measure the latency of a function by
> instrumenting its call site.  Basically, I find the call instruction,
> and insert a kprobe with a pre-handler and post-handler at that point.
> The pre-handler measures the latency (reads the TSC counter).  The
> post-handler measures the latency again, and subtracts the value that
> was read in the pre-handler to compute the total latency of the called
> function.

This sounds ok...

> So to measure the latency of foo(), I basically want kprobes to do this:
> pre_handler();
> foo();
> post_handler();
> 
> The problem is that the latencies that I am getting are consistently low
> (~10,000 cycles).  When I manually instrument the functions, the latency
> is about 20,000,000 cycles.  Clearly something is not right here.

Is foo() called from too many different places? If so, are you
interested with only the invocation of foo() from a specific callsite?
 
> Is this a known issue?  Instead of using the post-handler, I can try to
> add a kprobe to the following instruction with a pre-handler.  I was
> just curious if there was something fundamentally wrong with the
> approach I took, or maybe a bug that you should be made aware of.

I am not too sure... single-stepping a "call" instruction from a
different memory location (single-stepping out of line) requires some
fixups and kprobes handles such fixups just fine (see resume_execution()
in arch/<arch>/kernel/kprobes.c)

You could try a a couple of approaches for starters.

a. As you mention above, a kprobe on the function invocation and the
other on the instruction following the call; both need just pre_handlers. 

b.
- Insert a kprobe and a kretprobe on foo()
- The kprobe needs to have only a pre_handler that'll measure the latency
- A similar handler for the kretprobe handler can measure the latency
again and their difference will give you foo()'s latency.

<b> though will require you to do some housekeeping in case foo() is
reentrant to track which return instance corresponds to which call.

Ananth

PS: There was a thought of providing a facility to run a handler at
function entry even when just a kretprobe is used. Maybe we need to
relook at that; it'd have been useful in this case.