On Tue, Sep 16, 2025 at 11:52:20AM -0400, Steven Rostedt wrote:
> On Mon, 15 Sep 2025 18:19:53 -0700
> Kalesh Singh <[email protected]> wrote:
> 
> 
> > Hi Steve,
> > 
> > Thanks for the comments and suggestion you are right we can use bpf to
> > get the comm. There is nothing special about this trace event.  I will
> > drop comm in the next revision.
> > 
> > The reason I did the task_struct parameter (current): I believe there
> > is a limitation that we must  specify at least 1 parameter to the
> > TRACE_EVENT()  PROTO and ARGS macros.
> 
> OK, then this is another issue. We don't want tracepoint "markers".
> Each tracepoint can take up to 5K in memory due to the code it
> generates and the meta data to control it.
> 
> For something like that, we highly recommend dynamic probes (fprobes,
> kprobes, etc).
> 
> The only purpose of a static tracepoint is to get data within a
> function that is too difficult to get via a probe. It should never be
> used as a trigger where its purpose is "we hit this path".
>

Isn't the usual problem with that approach, that of static functions/static
inline functions? I was tracing through a problem a few months ago, and man
I really did think "wouldn't it be nice to have a tracepoint instead of fishing
around for kprobe spots".

Not that I particularly think a tracepoint is super worth it in this case, but,
y'know.

-- 
Pedro

Reply via email to