Re: [PATCH v4] [RFC] trace: Add kprobe on tracepoint

From: Masami Hiramatsu
Date: Thu Aug 12 2021 - 11:07:04 EST

On Thu, 12 Aug 2021 09:44:39 -0400
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Thu, 12 Aug 2021 20:31:10 +0900
> Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> > > Yes, anyway we need a way to find loops on histogram/eprobe at last.
> >
> > BTW, what about using similar machanism of "current_kprobe()" to detect
> > the reccursion? As an easy way, prepare a static per-cpu pointer which sets
> > the current eprobe and if the eprobe handler detects that is already set,
> > it may warn (or silently ignore) and reject it.
> > (Of course it is better to detect the loop when user sets the hist-trigger
> > by reverse link)
> Thinking more about this, I believe there is a use case for synthetic
> event on a eprobe. Basically:
> normal_event -> eprobe (extracts struct data into $dat) -> onmax($dat) -> synthetic event
> But I can not come up with any use case of:
> eprobe -> synthetic event -> eprobe
> or
> synthetic event -> eprobe -> synthetic event
> That's because once you have an eprobe, you can extract what you want,
> and once you have that synthetic event, you can get the data you want.
> Maybe we should prevent the above and allow one eprobe on a synthetic
> event and one synthetic event on an eprobe.
> Or just don't prevent it at all, and let the user shoot themselves in
> the foot ;-)
> The more I think about this, I'm thinking we just let them shoot
> themselves if they want to.

I agree. Or, at least we can prevent the loop at runtime as I said.
BTW, does synthetic event itself detect and prevent loops? I think
the key point is always synthetic event, so if the loop detector
is implemented, it should be done on the synthetic event.

> But I still agree that eprobes should not be attached to kprobes or
> uprobes directly (although they may be able to be attached to a
> synthetic event that is attached to one!)


Thank you,

> -- Steve

Masami Hiramatsu <mhiramat@xxxxxxxxxx>