Hi Steve,
On Fri, Jan 13, 2017 at 11:16:18AM -0500, Steven Rostedt wrote:
> On Fri, 13 Jan 2017 13:22:43 +0900
> Namhyung Kim <namhy...@kernel.org> wrote:
>
> > It's currently fixed to 32 and it ignores when user gives a pattern
> > which match to functions more than the size. So filtering like all
> > system calls or many functions with common prefix cannot be set all.
> > Not sure this is right though.
>
> Yes it's small, and there's a reason for it. So I'm giving a
> conditional nack to the patch.
> >
> > This patch adds 'graph_filter_size' file in the tracefs to adjust the
> > size. It can be changed only if the current tracer is not set.
> >
> > Signed-off-by: Namhyung Kim <namhy...@kernel.org>
>
> The condition is, we need to fix ftrace_graph_addr() first.
>
> for (i = 0; i < ftrace_graph_count; i++) {
>
> That gets called at every function being traced. See where I'm heading
> with that? ;-)
>
> We need to create a hash table or binary search first and make that
> function handle a large ftrace_graph_count before implementing your
> patch. Remove the linear search, replace it with either a binary search
> or a hash. But an O(n) algorithm at every function call is out of the
> question.
Fair enough. I'll try to add a hash table then.
But I'm not sure how to synchronize hash table manipulations. It
seems synchronize_sched() is not good enough for function graph
tracer, right? So I limited changing filter size only when no tracer
is used, but is it ok to have the limitation when adding or removing
an entry to/from the table? If not, what can I do?
Thanks,
Namhyung
