From mboxrd@z Thu Jan  1 00:00:00 1970
From: will.deacon@arm.com (Will Deacon)
Date: Tue, 18 Aug 2015 09:21:36 +0100
Subject: [RFC v2 1/4] ftrace: allow arch-specific check_stack()
In-Reply-To: <55D17A04.7020800@linaro.org>
References: <1438674249-3447-1-git-send-email-takahiro.akashi@linaro.org>
 <1438674249-3447-2-git-send-email-takahiro.akashi@linaro.org>
 <20150811170321.GB29880@arm.com> <55D17A04.7020800@linaro.org>
Message-ID: <20150818082136.GA10301@arm.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Mon, Aug 17, 2015 at 07:07:00AM +0100, AKASHI Takahiro wrote:
> On 08/12/2015 02:03 AM, Will Deacon wrote:
> > On Tue, Aug 04, 2015 at 08:44:06AM +0100, AKASHI Takahiro wrote:
> >> A stack frame pointer may be used in a different way depending on
> >> cpu architecture. Thus it is not always appropriate to slurp the stack
> >> contents, as currently done in check_stack(), in order to calcurate
> >> a stack index (height) at a given function call. At least not on arm64.
> >>
> >> This patch extract potentially arch-specific code from check_stack()
> >> and puts it into a new arch_check_stack(), which is declared as weak.
> >> So we will be able to add arch-specific and most efficient way of
> >> stack traversing Later.
> >>
> >> Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
> >
> > If arm64 is the only architecture behaving differently, then I'm happy
> > to reconsider the fix to unwind_frame that we merged in e306dfd06fcb
> > ("ARM64: unwind: Fix PC calculation"). I'd have thought any architecture
> > with a branch-and-link instruction would potentially have the same issue,
> > so we could just be fixing things in the wrong place if ftrace works
> > everywhere else.
> 
> I'm not the right person to answer for other architectures (and ftrace
> behavior on them.) The only thing I know is that current ftrace stack tracer
> works correctly only if the addresses stored and found on stack match to
> the ones returned by save_stack_trace().
> 
> Anyway, the fix above is not the only reason that I want to introduce arch-specific
> arch_check_stack(). Other issues to fix include
>    - combined case of stack tracer and function graph tracer (common across arch's)
>    - exception entries (as I'm trying to address in RFC 4/4)
>    - in-accurate stack size (for each function, my current fix is not perfect though.)

Ok, if you have other reasons for the callback, then fine. I just didn't
want you to think that you had to work around e306dfd06fcb at all costs.

Will