Re: [PATCH 1/2] powerpc/ftrace: support CONFIG_FUNCTION_GRAPH_RETVAL

[Aditya Bodkhe <adityab1@linux.ibm.com>]

On 03/06/25 2:22 pm, Christophe Leroy wrote:
>
>
> Le 28/05/2025 à 15:48, Aditya Bodkhe a écrit :
>> [Vous ne recevez pas souvent de courriers de adityab1@linux.ibm.com. 
>> Découvrez pourquoi ceci est important à 
>> https://aka.ms/LearnAboutSenderIdentification ]
>>
>> From: Aditya Bodkhe <aditya.b1@linux.ibm.com>
>>
>> commit a1be9ccc57f0 ("function_graph: Support recording and printing the
>> return value of function") introduced support for function graph return
>> value tracing.
>>
>> Additionally, commit a3ed4157b7d8 ("fgraph: Replace fgraph_ret_regs with
>> ftrace_regs") further refactored and optimized the implementation,
>> making `struct fgraph_ret_regs` unnecessary.
>>
>> This patch enables the above modifications for powerpc64, ensuring that
>> function graph return value tracing is available on this architecture.
>
> Why only powerpc64 ?
>
> I see nothing specific to powerpc64 in your patch, will it work on 
> powerpc32 too ?
>
That's a oversight on my part . This patch necessary changes for 
powerpc32 bit and tested this feature to be working as expected on the same.
>>
>> After this patch, v6.14+ kernel can also be built with FPROBE on powerpc
>> but there are a few other build and runtime dependencies for FPROBE to
>> work properly. The next patch addresses them.
>>
>> Signed-off-by: Aditya Bodkhe <aditya.b1@linux.ibm.com>
>> ---
>>   arch/powerpc/Kconfig                     |  1 +
>>   arch/powerpc/include/asm/ftrace.h        | 15 +++++++++
>>   arch/powerpc/kernel/trace/ftrace_entry.S | 41 ++++++++++++++----------
>>   3 files changed, 40 insertions(+), 17 deletions(-)
>>
>> diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
>> index c3e0cc83f120..9163521bc4b9 100644
>> --- a/arch/powerpc/Kconfig
>> +++ b/arch/powerpc/Kconfig
>> @@ -250,6 +250,7 @@ config PPC
>>          select HAVE_FUNCTION_ARG_ACCESS_API
>>          select HAVE_FUNCTION_DESCRIPTORS        if PPC64_ELF_ABI_V1
>>          select HAVE_FUNCTION_ERROR_INJECTION
>> +       select HAVE_FUNCTION_GRAPH_FREGS
>>          select HAVE_FUNCTION_GRAPH_TRACER
>>          select HAVE_FUNCTION_TRACER             if !COMPILE_TEST && 
>> (PPC64 || (PPC32 && CC_IS_GCC))
>>          select HAVE_GCC_PLUGINS                 if GCC_VERSION >= 
>> 50200   # plugin support on gcc <= 5.1 is buggy on PPC
>> diff --git a/arch/powerpc/include/asm/ftrace.h 
>> b/arch/powerpc/include/asm/ftrace.h
>> index 82da7c7a1d12..6ffc9c9cf4e3 100644
>> --- a/arch/powerpc/include/asm/ftrace.h
>> +++ b/arch/powerpc/include/asm/ftrace.h
>> @@ -50,6 +50,21 @@ static __always_inline struct pt_regs 
>> *arch_ftrace_get_regs(struct ftrace_regs *
>>                  asm volatile("mfmsr %0" : "=r" ((_regs)->msr)); \
>>          } while (0)
>>
>> +#undef ftrace_regs_get_return_value
>> +static __always_inline unsigned long
>> +ftrace_regs_get_return_value(const struct ftrace_regs *fregs)
>> +{
>> +       return arch_ftrace_regs(fregs)->regs.gpr[3];
>> +}
>> +#define ftrace_regs_get_return_value ftrace_regs_get_return_value
>> +
>> +#undef ftrace_regs_get_frame_pointer
>> +static __always_inline unsigned long
>> +ftrace_regs_get_frame_pointer(const struct ftrace_regs *fregs)
>> +{
>> +       return arch_ftrace_regs(fregs)->regs.gpr[1];
>> +}
>> +
> Why unset and redefine ftrace_regs_get_return_value() and 
> ftrace_regs_get_frame_pointer() ? Please explain why the default ones 
> can't be used on powerpc.
>
Because for powerpc currently for ftrace_regs_get_return_value() this 
function regs_return_value() is getting called and it's implementation 
returns -regs.gpr[3] if is_syscall_success() returns false . This was 
causing fails in ftrace selftests .

Also for ftrace_regs_get_frame_pointer() current implementation was 
always returning 0 hence we had to change that too .

>>   static __always_inline void
>>   ftrace_regs_set_instruction_pointer(struct ftrace_regs *fregs,
>>                                      unsigned long ip)
>> diff --git a/arch/powerpc/kernel/trace/ftrace_entry.S 
>> b/arch/powerpc/kernel/trace/ftrace_entry.S
>> index 3565c67fc638..eafbfb7584ed 100644
>> --- a/arch/powerpc/kernel/trace/ftrace_entry.S
>> +++ b/arch/powerpc/kernel/trace/ftrace_entry.S
>> @@ -409,23 +409,30 @@ EXPORT_SYMBOL(_mcount)
>>   _GLOBAL(return_to_handler)
>>          /* need to save return values */
>>   #ifdef CONFIG_PPC64
>> -       std     r4,  -32(r1)
>> -       std     r3,  -24(r1)
>> +       stdu    r1, -SWITCH_FRAME_SIZE(r1)
>> +       std     r4, GPR4(r1)
>> +       std     r3, GPR3(r1)
>> +  /* Save previous stack pointer (r1) */
>> +       addi    r3, r1, SWITCH_FRAME_SIZE
>> +       std     r3, GPR1(r1)
>>          /* save TOC */
>> -       std     r2,  -16(r1)
>> -       std     r31, -8(r1)
>> +       std     r2,  24(r1)
>> +       std     r31, 32(r1)
>>          mr      r31, r1
>> -       stdu    r1, -112(r1)
>> -
>> +  /* pass ftrace_regs/pt_regs to ftrace_return_to_handler */
>> +       addi    r3,  r1, STACK_INT_FRAME_REGS
>
> Some of the changes seems to only be renaming and should be done in a 
> cleanup/preparatory patch in order to only focus on real necessary 
> changes in this patch.
>
Now the way we are setting up the frame we cannot have negative offsets 
, hence we used offsets which are defined already.
>>          /*
>>           * We might be called from a module.
>>           * Switch to our TOC to run inside the core kernel.
>>           */
>>          LOAD_PACA_TOC()
>>   #else
>> -       stwu    r1, -16(r1)
>> -       stw     r3, 8(r1)
>> -       stw     r4, 12(r1)
>> +       stwu    r1, -SWITCH_FRAME_SIZE(r1)
> Why do we need such a big frame size just to save two registers ?
>
This Frame size is needed because we need pt_regs as part of frame , are 
you saying instead of SWITCH_FRAME_SIZE for 32 bit architecture we 
should just setup a frame with size of pt_regs because ftrace_regs size 
is same as pt_regs on ppc
>> +       stw     r4, GPR4(r1)
>> +       stw     r3, GPR3(r1)
>> +       addi    r3, r1, SWITCH_FRAME_SIZE
>> +       stw     r3, GPR1(r1)
> > +       addi    r3, r1, STACK_INT_FRAME_REGS
>
> Why is this needed ?
>
The function ftrace_return_to_handler takes pointer to pt_regs as 
arguement so we are setting that up in stack and passing that pointer.
>>   #endif
>>
>>          bl      ftrace_return_to_handler
>> @@ -435,15 +442,15 @@ _GLOBAL(return_to_handler)
>>          mtlr    r3
>>
>>   #ifdef CONFIG_PPC64
>> -       ld      r1, 0(r1)
>> -       ld      r4,  -32(r1)
>> -       ld      r3,  -24(r1)
>> -       ld      r2,  -16(r1)
>> -       ld      r31, -8(r1)
>> +       ld      r4,  GPR4(r1)
>> +       ld      r3,  GPR3(r1)
>> +       ld      r2,  24(r1)
>> +       ld      r31, 32(r1)
>> +       ld      r1,  0(r1)
>>   #else
>> -       lwz     r3, 8(r1)
>> -       lwz     r4, 12(r1)
>> -       addi    r1, r1, 16
>> +       lwz     r3, GPR3(r1)
>> +       lwz     r4, GPR4(r1)
>> +       addi    r1, r1, SWITCH_FRAME_SIZE
>>   #endif
>>
>>          /* Jump back to real return address */
>> -- 
>> 2.43.5
>>
>