Re: [RFC PATCH] tracing: tprobe-events: Register tracepoint when enable tprobe event

[Google]

On Thu, 1 May 2025 11:27:12 -0400
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Mon, 17 Mar 2025 17:03:16 +0900
> "Masami Hiramatsu (Google)" <mhiramat@xxxxxxxxxx> wrote:
> 
> > From: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > 
> > As same as fprobe, register tracepoint stub function only when enabling
> > tprobe events. The major changes are introducing a list of
> > tracepoint_user and its lock, and tprobe_event_module_nb, which is
> > another module notifier for module loading/unloading.  By spliting the
> > lock from event_mutex and a module notifier for trace_fprobe, it
> > solved AB-BA lock dependency issue between event_mutex and
> > tracepoint_module_list_mutex.
> > 
> > Signed-off-by: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > ---
> >  kernel/trace/trace_fprobe.c |  382 +++++++++++++++++++++++++------------------
> >  1 file changed, 218 insertions(+), 164 deletions(-)
> 
> Is this patch still needed? It doesn't apply cleanly.

This is a kind of performance optimization. Without this patch,
tprobe always registers a stub function to the tracepoint. This
*may* introduce an overhead, and it depends on how frequently the
tracepoint is used. But I guess it is not noticable unless we
create so many tprobes because the stub function is just a 'ret'.

Thus this is not so hurry. Anyway, I will update it for for-next
because this can clean up __tracepoint_probe_module_cb() logic too.

Thank you,

> 
> -- Steve


-- 
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>