Re: [PATCH 1/3] [BUGFIX] tracing: Fix double free of event_trigger_data

[Tom Zanussi]

Hi Steve,

On 7/24/2018 4:30 PM, Steven Rostedt wrote:
> On Tue, 24 Jul 2018 16:49:59 -0400
> Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> > > Hmm it seems we should review the register_trigger() implementation.
> > > It should return the return value of trace_event_trigger_enable_disable(),
> > > shouldn't it?
> > >   
> >
> > Yeah, that's not done well. I'll fix it up.
> >
> > Thanks for pointing it out.
>
> Tom,
>
> register_trigger() is messed up. I should have caught this when it was
> first submitted, but I'm totally confused. The comments don't match the
> code.
>
> First we have this:
>
> 	ret = cmd_ops->reg(glob, trigger_ops, 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.
> 	 */
>
> Which looks to be total BS.

Yes, it is BS in the case of event triggers.  This was taken from the 
ftrace function trigger code, where it does make sense.  I think I left 
it in thinking the code would at some point later converge.

> As we have this:
>
> /**
>   * register_trigger - Generic event_command @reg implementation
>   * @glob: The raw string used to register the trigger
>   * @ops: The trigger ops associated with the trigger
>   * @data: Trigger-specific data to associate with the trigger
>   * @file: The trace_event_file associated with the event
>   *
>   * Common implementation for event trigger registration.
>   *
>   * Usually used directly as the @reg method in event command
>   * implementations.
>   *
>   * Return: 0 on success, errno otherwise

And this is how it should work.

>   */
> static int register_trigger(char *glob, struct event_trigger_ops *ops,
> 			    struct event_trigger_data *data,
> 			    struct trace_event_file *file)
> {
> 	struct event_trigger_data *test;
> 	int ret = 0;
>
> 	list_for_each_entry_rcu(test, &file->triggers, list) {
> 		if (test->cmd_ops->trigger_type == data->cmd_ops->trigger_type) {
> 			ret = -EEXIST;
> 			goto out;
> 		}
> 	}
>
> 	if (data->ops->init) {
> 		ret = data->ops->init(data->ops, data);
> 		if (ret < 0)
> 			goto out;
> 	}
>
> 	list_add_rcu(&data->list, &file->triggers);
> 	ret++;
>
> 	update_cond_flag(file);
> 	if (trace_event_trigger_enable_disable(file, 1) < 0) {
> 		list_del_rcu(&data->list);
> 		update_cond_flag(file);
> 		ret--;
> 	}
> out:
> 	return ret;
> }
>
> Where the comment is total wrong. It doesn't return 0 on success, it
> returns 1. And if trace_event_trigger_enable_disable() fails it returns
> zero.
>
> And that can fail with the call->class->reg() return value, which could
> fail for various strange reasons. I don't know why we would want to
> return 0 when it fails?
>
> I don't see where ->reg() would return anything but 1 on success. Maybe
> I'm missing something. I'll look some more, but I'm thinking of changing
> ->reg() to return zero on all success, and negative on all errors and
> just check those results.

Right, in the case of event triggers, we only register one at a time, 
whereas with the trace function triggers, with globbing multiple 
functions can register triggers at the same time, so it makes sense 
there to have reg() return a count and the more convoluted error handling.

So I agree, simplifying things here by using the standard error handling 
would be an improvement.

Tom

> -- Steve