[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20200622103414.af303c4d4b0dad1c9d7262a3@kernel.org>
Date: Mon, 22 Jun 2020 10:34:14 +0900
From: Masami Hiramatsu <mhiramat@...nel.org>
To: Ming Lei <ming.lei@...hat.com>
Cc: Steven Rostedt <rostedt@...dmis.org>,
Ming Lei <tom.leiming@...il.com>,
"Naveen N. Rao" <naveen.n.rao@...ux.ibm.com>,
Anil S Keshavamurthy <anil.s.keshavamurthy@...el.com>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
"David S. Miller" <davem@...emloft.net>,
linux-block <linux-block@...r.kernel.org>
Subject: Re: kprobe: __blkdev_put probe is missed
On Mon, 22 Jun 2020 08:27:53 +0800
Ming Lei <ming.lei@...hat.com> wrote:
> I mean it isn't from user's viewpoint, and the binary code is usually a
> black box for final kprobe user.
>
> IMO, all your and Steven's input are just from kprobe/trace developer's viewpoint.
> Can you think about the issue from kprobe real/final user?
>
> Trace is very useful tools to observe system internal, and people often
> relies on trace to understand system. However, missed probe often causes
> trouble for us to understand the system correctly.
Agreed. However, since kprobes related tracing tools are layered
to provide different features (e.g. kprobes abstructs sw breakpoint,
ftrace kprobe-events provides a minimum CUI, and perf-probe provides
binary analysis, etc.), this issue should be solved by user-level
binary analysis layer. (it is not good idea to analyze the optimized
code in kernel)
> > > 2) from implementation view, I understand exception should be trapped
> > > on the entry of __blkdev_put(), looks it isn't done.
> >
> > No, it is correctly trapped the function entry address. The problem is
> > that the gcc optimized the nested function call into jump to the
> > beginning of function body (skip prologue).
> >
> > Usually, a function is compiled as below
> >
> > func() (1) the entry address (func:)
> > { (2) the function prologue (setup stackframe)
> > int a (3) the beginning of function body
> > ...
> > func() (4) the nested function call
> >
> > And in this case, the gcc optimized (4) into jump to (3) instead of
> > actual function call instruction. Thus, for the nested case (1) and
> > (2) are skipped.
> > IOW, the code flow becomes
> > (1)->(2)->(3)->(4)->(3)
> > instead of
> > (1)->(2)->(3)->(4)->(1)->(2)->(3)
> >
> > In this case, if we put a probe on (1) or (2), those are disappeared
> > in the nested call. Thus if you put a probe on (3) ('perf probe __blkdev_put:2')
> > you'll see the event twice.
>
> Thanks for your explanation.
>
> Can you kprobe guys improve the implementation for covering this case?
> For example, put probe on 3) in case the above situation is recognized.
OK, let me try to fix this in perf-probe since that is the simplest
binary analysis part in user-space.
Thank you,
--
Masami Hiramatsu <mhiramat@...nel.org>
Powered by blists - more mailing lists