Re: [PATCH] arm: ftrace: Adds support for CONFIG_DYNAMIC_FTRACE_WITH_REGS

[Jean-Jacques Hiblot]

2017-01-12 15:30 GMT+01:00 Jean-Jacques Hiblot <jjhiblot@xxxxxxxxxxxxxxx>:
> 2017-01-12 1:19 GMT+01:00 Abel Vesa <abelvesa@xxxxxxxxx>:
>> On Tue, Jan 10, 2017 at 04:51:12PM +0100, Petr Mladek wrote:
>>> On Thu 2016-12-08 22:46:55, Abel Vesa wrote:
>>> > On Thu, Dec 08, 2016 at 09:46:35PM +0000, Abel Vesa wrote:
>>> > > From: Jean-Jacques Hiblot <jjhiblot@xxxxxxxxxxxxxxx>
>>> > >
>>> > > From: Jean-Jacques Hiblot <jjhiblot@xxxxxxxxxxxxxxx>
>>> >
>>> > >From statement twice in the commit message. Will resend.
>>> > >
>>> > > The DYNAMIC_FTRACE_WITH_REGS configuration makes it possible for a ftrace
>>> > > operation to specify if registers need to saved/restored by the ftrace handler.
>>> > > This is needed by kgraft and possibly other ftrace-based tools, and the ARM
>>> > > architecture is currently lacking this feature. It would also be the first step
>>> > > to support the "Kprobes-on-ftrace" optimization on ARM.
>>> > >
>>> > > This patch introduces a new ftrace handler that stores the registers on the
>>> > > stack before calling the next stage. The registers are restored from the stack
>>> > > before going back to the instrumented function.
>>> > >
>>> > > A side-effect of this patch is to activate the support for ftrace_modify_call()
>>> > > as it defines ARCH_SUPPORTS_FTRACE_OPS for the ARM architecture
>>> > >
>>> > > Signed-off-by: Jean-Jacques Hiblot <jjhiblot@xxxxxxxxxxxxxxx>
>>> > > Signed-off-by: Abel Vesa <abelvesa@xxxxxxxxx>
>>> > > ---
>>> > >  arch/arm/Kconfig               |  2 ++
>>> > >  arch/arm/include/asm/ftrace.h  |  4 +++
>>> > >  arch/arm/kernel/entry-ftrace.S | 78 ++++++++++++++++++++++++++++++++++++++++++
>>> > >  arch/arm/kernel/ftrace.c       | 33 ++++++++++++++++++
>>> > >  4 files changed, 117 insertions(+)
>>> > >
>>> > > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>> > > index b5d529f..87f1a9f 100644
>>> > > --- a/arch/arm/Kconfig
>>> > > +++ b/arch/arm/Kconfig
>>> > > @@ -50,6 +50,7 @@ config ARM
>>> > >   select HAVE_DMA_API_DEBUG
>>> > >   select HAVE_DMA_CONTIGUOUS if MMU
>>> > >   select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL) && !CPU_ENDIAN_BE32 && MMU
>>> > > + select HAVE_DYNAMIC_FTRACE_WITH_REGS if HAVE_DYNAMIC_FTRACE
>>> > >   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)
>>> > > @@ -90,6 +91,7 @@ config ARM
>>> > >   select PERF_USE_VMALLOC
>>> > >   select RTC_LIB
>>> > >   select SYS_SUPPORTS_APM_EMULATION
>>> > > + select FRAME_POINTER if DYNAMIC_FTRACE_WITH_REGS && FUNCTION_GRAPH_TRACER
>> Hi Petr,
>>>
>>> FRAME_POINTER is not for free. It takes space on the stack. Also there
>>> is a performance penalty. Do we really need to depend on it? If so,
>>> it might be worth a note in the commit message.
>>
>
> FRAME_POINTER is not needed. the dependency is wrong and should be removed.
> The code must be modified to not use fp register:
>
> --- a/arch/arm/kernel/entry-ftrace.S
> +++ b/arch/arm/kernel/entry-ftrace.S
> @@ -130,7 +130,8 @@ ftrace_graph_regs_call:
>  #ifdef CONFIG_FUNCTION_GRAPH_TRACER
>  .macro __ftrace_graph_regs_caller
>
> -       sub     r0, fp, #4                      @ lr of instrumented
> routine (parent)
> +       add     r0, sp, #64             @ r0 is now a pointer to lr of
> +                                       @ instrumented routine

I made some tests after sending this email. And it turns out that it
doesn't work if we change  "sub     r0, fp, #4" to "add     r0, sp,
#64 " here.
So it looks like there is a dependency on FRAME_POINTER after all.
Note that the same is true for  __ftrace_graph_caller. I don't know if
the 'graph' feature of ftrace requires intrinsically FRAME_POINTER but
it looks like it currently does on ARM (with or without register
saving)
I'll try to spend some time on the subject next week.

>
>         @ called from __ftrace_regs_caller
>         ldr     r1, [sp, #56]                   @ instrumented routine (func)
> @@ -139,8 +140,9 @@ ftrace_graph_regs_call:
>         mov     r2, fp                          @ frame pointer
>         bl      prepare_ftrace_return
>
> -       ldr     lr, [fp, #-4]                   @ restore lr from the stack
> -       ldmia   sp, {r0-r11, ip, sp}            @ restore r0 through sp
> +       ldr     lr, [sp, #64]           @ get the previous LR value from stack
> +       ldmia   sp, {r0-r11, ip, sp}    @ pop the saved registers INCLUDING
> +                                       @ the stack pointer
>         ret     ip
>  .endm
>  #endif
>
>
> Jean-Jacques
>
>
>> I was trying to create my own patch when I found this work done by
>> Jean-Jacques, so I haven't looked specifically for the FRAME_POINTER
>> part. I looked now at it and you seem to be right, FRAME_POINTER is
>> not needed.
>>
>> I will get rid of the FRAME_POINTER part, change the authorship and
>> send it again in the following days.
>>>
>>> I made only a quick look at the patch. It looks reasonable. But I do
>>> not have enough knowledge about the arm architecture, assembly, and
>>> ftrace-specifics. Also I cannot test it easily. So issues might
>>> be hidden to my eyes.
>>>
>>> Best Regards,
>>> Petr
>> Thanks,
>> Abel