Re: [RFC PATCH v2 12/31] function_graph: Have the instances use their own ftrace_ops for filtering

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

On Thu, 9 Nov 2023 21:18:48 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:

> On Fri, 10 Nov 2023 10:51:54 +0900
> Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:
> 
> > So this patch registers ftrace_ops for each fgraph_ops to ftrace.
> > This means that the ftrace_graph_func() will be called twice or more
> > on the same function.
> > Thus should I call ftrace_startup() once when the first fgraph_ops
> > is registered? 
> > No, it's not enough. Actually each fgraph_ops can have different filters.
> > We need to define a shared filter and combine new filters to one and
> > use it. We also need to do it when a fgraph is unregistered.
> > 
> > Is there any function which makes a new filter from two (or more) filters?
> 
> So I'm guessing that we need to have a fgraph_set_filter*() operations?
> 
> When one gets added, it needs to update the ftrace_ops to include the added
> functions. Or we need to have a way to create a new hash from all the
> registered fgraph_ops, and have that for the ftrace_ops. Then when it gets
> called, if it has more than one registered function, it needs to iterate
> over the list?

Yes, that is one option, update a global common hash and introduce a new
common ftrace function to run function_graph_enter().

Or, I think keep the current one but iterate ftrace_ops to callback the
function_graph_enter() with ftrace_ops. Then we can get appropriate
fgraph_ops. Ftrace push return trace can skip pushing if ret == return_to_handler.
(maybe this is better to reuse ftrace)

Thank you,

> 
> -- Steve
> 
> 
> > 
> > Or, maybe we can make the common callback to find the previous ret entry on
> > the ret_stack and reuse it. (In this case we don't need loop on each
> > fgraph_array entry)
> 


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>