Re: KPROBES: Instrumenting a function's call site

[Avishay Traeger <atraeger@cs.sunysb.edu>]

On Wed, 2007-09-26 at 22:57 +0530, Ananth N Mavinakayanahalli wrote:
> On Wed, Sep 26, 2007 at 12:09:35PM -0400, Avishay Traeger wrote:
> > On Wed, 2007-09-26 at 14:33 +0530, Ananth N Mavinakayanahalli wrote:
> > > What happens when the "call" is singlestepped is that the instruction
> > > pointer is moved to the call target. That explains the lower latency you
> > > are seeing. You'll need to do something along the lines I suggested in
> > > the earlier mail.
> > 
> > Can you please explain what you mean by this more clearly?  I'm not a
> > kprobes expert yet.  Specifically, using kprobes the way that I did,
> > what will the resulting code look like?  Also, what do you mean by
> > "singlestepped"?
> 
> If you single-step (regs->eflags | TF_MASK in i386) on a call instruction,
> you'll end up at the call target; ie., after the post_kprobe_handler()
> returns, the instruction pointer will point to the first instruction
> of foo().
> 
> Try printk()ing the instruction pointer(regs) after resume_execution()
> in the post_kprobe_handler() in your arch/<arch>/kernel/kprobes.c, you'll
> see what I mean.
> 
> And when I say singlestepped, I mean executing one instruction under the
> architecture specific single step enable flag - the "trap" flag for i386,
> the MSR_SE for powerpc, etc. Evidently, this'll mean single-stepping a
> single instruction.
> 
> Ananth

I see - thanks for all your prompt and helpful advice!

Avishay