Re: [for-next][PATCH 05/31] tracing: Have existing event_command.parse() implementations use helpers

[Tom Zanussi <zanussi@kernel.org>]

Hi Steve,

On Thu, 2022-01-13 at 16:20 -0500, Steven Rostedt wrote:
> On Thu, 13 Jan 2022 18:03:07 +0100
> Daniel Bristot de Oliveira <bristot@kernel.org> wrote:
> 
> > I did some debug, and found that the histogram is working. The
> > problem is that,
> > to read the histogram I pause it to have consistent data:
> > 
> > in tools/tracing/rtla/osnoise_hist.c:
> > osnoise_read_trace_hist() {
> >  [...]
> >         tracefs_hist_pause(tool->trace.inst, data->trace_hist);
> > 
> >         content = tracefs_event_file_read(tool->trace.inst,
> > "osnoise",
> >                                           "sample_threshold",
> >                                           "hist", NULL);
> >  [...]
> > }
> > 
> > and, as far as I got, after this patch, pausing the histogram makes
> > it to clear
> > up. If I comment the "tracefs_hist_pause" line, "rtla osnoise hist"
> > start
> > working back again.
> > 
> > Thoughts?
> 
> This is all messed up. I'm removing this patch completely.
> 
> Tom, can you fix this. The issue is that it's putting too much policy
> into
> the helper functions, which is big no no.
> 
> Specifically, we have:
> 
> int event_trigger_register(struct event_command *cmd_ops,
> 			   struct trace_event_file *file,
> 			   char *glob,
> 			   char *cmd,
> 			   char *param,
> 			   struct event_trigger_data *trigger_data,
> 			   int *n_registered)
> {
> 	int ret;
> 
> 	if (n_registered)
> 		*n_registered = 0;
> 
> 	ret = cmd_ops->reg(glob, trigger_data, file);
> 	/*
> 	 * The above returns on success the # of functions enabled,
> 	 * but if it didn't find any functions it returns zero.
> 	 * Consider no functions a failure too.
> 	 */
> 	if (!ret) {
> 		cmd_ops->unreg(glob, trigger_data, file);
> 		ret = -ENOENT;
> 	} else if (ret > 0) {
> 		if (n_registered)
> 			*n_registered = ret;
> 		/* Just return zero, not the number of enabled
> functions */
> 		ret = 0;
> 	}
> 
> 	return ret;
> }
> 
> 
> And in the case of pause, this *will* have ret = 0 on return. And
> what
> happens is that it removes the trigger completely.
> 
> Look at the code in the histogram on the return:
> 
> 	ret = event_trigger_register(cmd_ops, file, glob, cmd, param,
> trigger_data, &n_registered);
> 	if (ret < 0)
> 		goto out_free;
> 	if ((ret == 0) && (n_registered == 0)) {
> 		if (!(attrs->pause || attrs->cont || attrs->clear))
> 			ret = -ENOENT;
> 		goto out_free;
> 	}
> 
> It checks for 0 and 0 and only errors if it's not pause, cont, or
> clear.
> Hence, all three are now broken due to this patch.
> 
> I will not be adding this to this merge window.

Yes, you're right, event_trigger_register() is trying to do a little
too much, and shouldn't be doing the unreg().

Thanks for finding and figuring that out.  I'll fix this and send a new
version after the merge window.

Tom

> 
> -- Steve