Message-ID: <20190414153704.7xlyluams3rl2mvo@shell.armlinux.org.uk>
Date:   Sun, 14 Apr 2019 16:37:04 +0100
From:   Russell King - ARM Linux admin <linux@...linux.org.uk>
To:     Masami Hiramatsu <mhiramat@...nel.org>
Cc:     Steven Rostedt <rostedt@...dmis.org>,
        Stefan Agner <stefan@...er.ch>, Arnd Bergmann <arnd@...db.de>,
        Masahiro Yamada <yamada.masahiro@...ionext.com>,
        linux-arm-kernel@...ts.infradead.org, naresh.kamboju@...aro.org,
        linux-kernel@...r.kernel.org
Subject: Re: Kconfig dependency issue on function-graph tracer and frame
 pointer on arm

On Sun, Apr 14, 2019 at 11:52:38PM +0900, Masami Hiramatsu wrote:
> On Sun, 14 Apr 2019 14:34:58 +0100
> Russell King - ARM Linux admin <linux@...linux.org.uk> wrote:
> 
> > On Sun, Apr 14, 2019 at 07:47:05PM +0900, Masami Hiramatsu wrote:
> > > Hello,
> > > 
> > > Recently, Naresh reported that the function-graph tracer on the latest
> > > kernel crashes on arm. I could reproduce it and bisected. I finally found
> > > the commit f9b58e8c7d03 ("ARM: 8800/1: use choice for kernel unwinders")
> > > was the first bad commit.
> > 
> > I don't think that littering the rest of the kernel Kconfig with ARM
> > specific stuff is really a viable solution to this.
> > 
> > If we examine the current situation, we have:
> > 
> > - THUMB2_KERNEL selecting ARM_UNWIND when enabled.
> > - UNWINDER_FRAME_POINTER disabled if THUMB2_KERNEL is enabled, provided
> >   we're not using Clang.  This leaves UNWINDER_ARM as the only choice,
> >   which also selects ARM_UNWIND.
> > - The default choice is dependent on the settings of AEABI and
> >   FUNCTION_GRAPH_TRACER.
> > - HAVE_FUNCTION_GRAPH_TRACER is disabled if THUMB2_KERNEL is enabled.
> > 
> > which seems to be _way_ too messy.
> > 
> > Looking back before this commit, the function graph tracer never had a
> > strong dependence on frame pointers being enabled in the kernel, but it
> > seems the code relies upon them (see ftrace_return_to_handler() in
> > kernel/trace/ and return_to_handler in arch/arm/kernel/entry-frace.S).
> > There is also the __ftrace_graph_caller macro which seems to rely on it.
> 
> Yes, so I think similar bug is hiding in other LTS kernels. It should
> have a dependency to FRAME_POINTER on arm.
> 
> > Since Clang does not support frame pointers, we shouldn't even offer
> > the function graph tracer for Clang compilers, so let's do that with
> > the first hunk of the patch below.
> > 
> > The subsequent hunks remove the defaulting of the choice according to
> > the function graph tracer - this is not a "hint" where the user can
> > still choose either option irrespective of the state of the function
> > graph tracer.  They should only be able to select the frame pointer
> > option in that case.
> 
> Agreed. Using default for making dependency is wrong.
> 
> > 
> > Another way forward would be for someone to put the work in to making
> > the function graph tracer work without frame pointers.
> 
> Yes, we eventually need that. But for fixing current released kernel
> (this bug is in v5.0 series), I think Kconfig fix is needed.
> 
> > 
> > So, how about this:
> > 
> > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> > index 850b4805e2d1..9aed25a6019b 100644
> > --- a/arch/arm/Kconfig
> > +++ b/arch/arm/Kconfig
> > @@ -73,7 +73,7 @@ config ARM
> >  	select HAVE_EFFICIENT_UNALIGNED_ACCESS if (CPU_V6 || CPU_V6K || CPU_V7) && MMU
> >  	select HAVE_EXIT_THREAD
> >  	select HAVE_FTRACE_MCOUNT_RECORD if !XIP_KERNEL
> > -	select HAVE_FUNCTION_GRAPH_TRACER if !THUMB2_KERNEL
> > +	select HAVE_FUNCTION_GRAPH_TRACER if !THUMB2_KERNEL && !CC_IS_CLANG
> >  	select HAVE_FUNCTION_TRACER if !XIP_KERNEL
> >  	select HAVE_GCC_PLUGINS
> >  	select HAVE_HW_BREAKPOINT if PERF_EVENTS && (CPU_V6 || CPU_V6K || CPU_V7)
> > diff --git a/arch/arm/Kconfig.debug b/arch/arm/Kconfig.debug
> > index 6d6e0330930b..e388af4594a6 100644
> > --- a/arch/arm/Kconfig.debug
> > +++ b/arch/arm/Kconfig.debug
> > @@ -47,8 +47,8 @@ config DEBUG_WX
> >  
> >  choice
> >  	prompt "Choose kernel unwinder"
> > -	default UNWINDER_ARM if AEABI && !FUNCTION_GRAPH_TRACER
> > -	default UNWINDER_FRAME_POINTER if !AEABI || FUNCTION_GRAPH_TRACER
> > +	default UNWINDER_ARM if AEABI
> > +	default UNWINDER_FRAME_POINTER if !AEABI
> 
> If UNWINDER_ARM depends on ARM EABI, would we really need this "if !AEABI"?
> (I doubt we need these default...)
> 
> >  	help
> >  	  This determines which method will be used for unwinding kernel stack
> >  	  traces for panics, oopses, bugs, warnings, perf, /proc/<pid>/stack,
> > @@ -65,7 +65,7 @@ config UNWINDER_FRAME_POINTER
> >  
> >  config UNWINDER_ARM
> >  	bool "ARM EABI stack unwinder"
> > -	depends on AEABI
> > +	depends on AEABI && !FUNCTION_GRAPH_TRACER
> 
> Hmm, AFAIK, FUNCTION_GRAPH_TRACER only depends on FRAME_POINTER, but not
> UNWINDER_FRAME_POINTER. So can user still choose UNWINDER_ARM even if
> FUNCTION_GRAPH_TRACER=y ? (Of course that may not be a meaningful option)

The UNWINDER_* options do not control anything except whether
FRAME_POINTER is enabled or not.  FRAME_POINTER controls not only
whether we build with frame pointers, but also how we unwind.
If both ARM_UNWIND and FRAME_POINTER are set, the kernel will
fail to link due to a multiple definition of unwind_frame().

The UNWINDER_* symbols were added in the commit you referenced
merely to select which of ARM_UNWIND or FRAME_POINTER are
enabled.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up

Powered by blists - more mailing lists