Re: Is it possible to trace events and its call stack?

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Mon, Jan 16, 2017 at 09:26:29PM +0900, Masami Hiramatsu escreveu:
> On Mon, 16 Jan 2017 16:48:57 +0800
> Qu Wenruo <quwenruo@cn.fujitsu.com> wrote:
> 
> > 
> > 
> > At 01/16/2017 10:55 AM, Masami Hiramatsu wrote:
> > > On Thu, 12 Jan 2017 15:49:08 +0800
> > > Qu Wenruo <quwenruo@cn.fujitsu.com> wrote:
> > >
> > >> Hi,
> > >>
> > >> Is it possible to use perf/ftrace to trace events and its call stack?
> > >>
> > >> [Background]
> > >> It's one structure in btrfs, btrfs_bio, I'm tracing for.
> > >> That structure is allocated and free somewhat frequently, and its size
> > >> is not fixed, so no SLAB/SLUB cache is used.
> > >>
> > >> I added trace events(or trace points, anyway, just in
> > >> include/trace/events/btrfs.h) to trace the allocation and freeing.
> > >> Which will output the pointer address of that structure, so I can pair
> > >> them, alone with other info.
> > >>
> > >> Things went well until, I found some structures are allocated but not
> > >> freed. (no corresponding trace point is triggered for given address).
> > >>
> > >> It's possible that btrfs just forget to free it, or btrfs is just
> > >> holding it for some purpose.
> > >> So kernel memleak detector won't catch the later one.
> > >>
> > >> That's to say along with the tracepoint data, I still need the call
> > >> stack of each calling, to determine the code who leak or hold the pointer.
> > >>
> > >> Is it possible to do it using perf or ftrace?
> > >
> > > If you are using ftrace, yes, you can enable stacktrace for each
> > > events by setting stacktrace event trigger as below;
> > >
> > > echo stacktrace > events/btrfs/<your event>/trigger
> > >
> > > Then ftrace will show the stacktrace data.
> > > See /sys/kernel/debug/tracing/README for more details. :)
> > 
> > That's great!
> > 
> > The most pure ftrace method!
> > 
> > I also found perf makes life easier compared to pure ftrace one :)
> 
> Should be :)
> 
> > Although after some search, I didn't find any equivalent of 
> > "function_graph" tracer, which is quite handy to handle small amount of 
> > calling time data.
> > 
> > Does it mean perf tool just doesn't support?
> 
> Namhyung had made it. I'm not sure why it is not merged.
> 
> https://lwn.net/Articles/570503/

/me trying to revive that patchset...

- Arnaldo