Re: ARM: config issue with ftrace function graph tracer

[Stefan Agner]

On 13.01.2019 17:14, Russell King - ARM Linux wrote:
> On Sun, Jan 13, 2019 at 04:41:52PM +0100, Stefan Agner wrote:
>> Hi,
>>
>> On 12.01.2019 02:01, Jeremy Fertic wrote:
>> > I'm having a problem with the ftrace function graph tracer on a 32 bit arm
>> > board (orangepi pc). A bisect points to the following commit:
>> >
>> > f9b58e8c7d03 ("ARM: 8800/1: use choice for kernel unwinders")
>> >
>> > Before this commit, if I use sunxi_defconfig and then menuconfig to enable
>> > FTRACE and FUNCTION_TRACER then the function graph tracer works. With this
>> > commit, and as of v5.0-rc1, doing the same as above results in a broken
>> > function graph tracer and often an oops as well. The commit introduces a
>> > choice group and it looks like it should default to UNWINDER_FRAME_POINTER
>> > if FUNCTION_GRAPH_TRACER is enabled. FUNCTION_GRAPH_TRACER is enabled by
>> > default when I enable FUNCTION_TRACER but this has no effect on the choice.
>> > The choice always defaults to the other option which is UNWINDER_ARM. If I
>> > manually choose UNWINDER_FRAME_POINTER then the function graph tracer works
>> > fine.
>>
>> The default selection is there, but this is made at "make
>> sunxi_defconfig" time. At this point FUNCTION_GRAPH_TRACER is not
>> enabled, hence Kconfig uses UNWIDER_ARM. However, when enabling the
>> FUNCTION_GRAPH_TRACER Kconfig will _not_ reconsider and switch enable
>> UNWINDER_FRAME_POINTER.
>>
>> Before that commit, when enabling FUNCTION_GRAPH_TRACER, we simply also
>> enabled FRAME_POINTER...
>>
>> I guess we need to make sure that FUNCTION_GRAPH_TRACER depends on the
>> UNWINDER_FRAME_POINTER choice. There is already a similar dependency
>> with THUMB2_KERNEL. We can cleanup that dependency since
>> UNWINDER_FRAME_POINTER already depends on !THUMB2_KERNEL.
>>
>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>> index 664e918e2624..a2ac65a8b2cc 100644
>> --- a/arch/arm/Kconfig
>> +++ b/arch/arm/Kconfig
>> @@ -69,7 +69,7 @@ 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 UNWINDER_FRAME_POINTER
>>         select HAVE_FUNCTION_TRACER if !XIP_KERNEL
>>         select HAVE_GCC_PLUGINS
>>         select HAVE_GENERIC_DMA_COHERENT
>> diff --git a/arch/arm/Kconfig.debug b/arch/arm/Kconfig.debug
>> index 6d6e0330930b..8341649fa71d 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
>>         help
>>           This determines which method will be used for unwinding kernel
>> stack
>>           traces for panics, oopses, bugs, warnings, perf,
>> /proc/<pid>/stack,
> 
> This looks rather horrid - the upshot of this means that the function
> tracer becomes unavailable on EABI unless you know that you must change
> the unwinder from its default.
> 
> Before the change to the choice statement, people could select the
> function graph tracer, and the correct unwinder would be selected
> for them.  That's knowledge that people never required before, and
> I think it's really quite unfair to require them to know this to use
> the function graph tracer.

It is not ideal I agree.

Searching for the symbol FUNCTION_GRAPH_TRACER in menuconfig immediately
shows what dependency are missing.

Also, we used to have the same situation with THUMB2_KERNEL already: You
had to know that you need to disable Thumb2 Kernel in order to see
FUNCTION_GRAPH_TRACER.


> 
> Maybe someone can put some effort into getting the function graph
> tracer working with non-framepointer kernels... but as the above
> currently stands, I really don't like the patch and I'd much rather
> revert the original change to fix this regression.

I am all for that effort. Using Thumb2 on Arm32 is also getting more
popular.

Is known what is exactly missing/what effort would be required?

--
Stefan