Re: [PATCH] trace: Add trace any kernel object

[Masami Hiramatsu]

On Thu, 21 Oct 2021 00:04:40 +0800
Huan Xie <xiehuan09@xxxxxxxxx> wrote:

> > > > For example:
> > > >
> > > > For the function bio_add_page, we can trace the first argument:
> > > >
> > > > int bio_add_page(struct bio *bio, struct page *page,
> > > >                             unsigned int len, unsigned int offset)
> > > >
> > > >     #cd /sys/kernel/debug/tracing
> > > >     #echo 1 > ./trace_object
> >
> > Hmm, so what happen if there are several events in parallel?
> >
> > > >     #echo 'p bio_add_page arg1=$arg1' > kprobe_events
> > > >     #echo 1 > ./events/kprobes/p_bio_add_page_0/enable
> >
> > what about adding a "objfilter" trigger action to update the object
> > list? e.g.
> >
> > echo objfilter:add:arg1:1 if comm == "sync" >> ./events/kprobes/p_bio_add_page_0/trigger
> >
> > This will "add the value of 'arg1' to objfilter 1 time if the process
> > name is sync".
> 
> Thanks. The trigger seems to work in combination with kprobe/uprobe/eprobe.
> 
> I have been trying to implement this feature in recent days, But my
> code does not look very elegant at the moment :-)

No problem! I can help you to brush up the coding :)

> It seems the result of 'cat  trace' should also need to be filtered,
> Because it will output the functions we don’t want to care about.

It actually depends on what object user wants to trace. e.g. network
packet trace will not be affected by 'cat trace'.
Anyway, we have "options/pause-on-trace" to stop tracing while
reading trace file. So user can choose it stop it or not. :)

> 
> I didn't expect this idea to be a relatively large project. :-)
> 

Because you have an exciting idea :)

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>