[RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Alexey Kardashevskiy]

For whatever reason LLVM does not allow LTO (Link Time Optimization) if
FTRACE_MCOUNT_USE_RECORDMCOUNT is enabled.

This allows disabling just this option instead of disabling all FTRACE
options.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---

Or disabling FTRACE is the right thing to do if HAS_LTO_CLANG=y?

Came from arch/Kconfig:

config HAS_LTO_CLANG
        def_bool y
        depends on CC_IS_CLANG && LD_IS_LLD && AS_IS_LLVM
        depends on $(success,$(NM) --help | head -n 1 | grep -qi llvm)
        depends on $(success,$(AR) --help | head -n 1 | grep -qi llvm)
        depends on ARCH_SUPPORTS_LTO_CLANG
        depends on !FTRACE_MCOUNT_USE_RECORDMCOUNT  <======
        depends on !KASAN || KASAN_HW_TAGS
        depends on !GCOV_KERNEL


---
 kernel/trace/Kconfig | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
index a5eb5e7fd624..87d82d2b0b0b 100644
--- a/kernel/trace/Kconfig
+++ b/kernel/trace/Kconfig
@@ -704,7 +704,8 @@ config FTRACE_MCOUNT_USE_OBJTOOL
 	depends on FTRACE_MCOUNT_RECORD
 
 config FTRACE_MCOUNT_USE_RECORDMCOUNT
-	def_bool y
+	bool "Enable FTRACE_MCOUNT_USE_RECORDMCOUNT"
+	default y
 	depends on !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY
 	depends on !FTRACE_MCOUNT_USE_CC
 	depends on !FTRACE_MCOUNT_USE_OBJTOOL
-- 
2.30.2

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Nick Desaulniers]

+ Sami, Kees

Do you guys remember why we don't support
FTRACE_MCOUNT_USE_RECORDMCOUNT with LTO?
https://patchwork.kernel.org/project/linux-kbuild/patch/20201013003203.4168817-12-samitolvanen@google.com/?

On Thu, Feb 10, 2022 at 5:43 PM Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>
> For whatever reason LLVM does not allow LTO (Link Time Optimization) if
> FTRACE_MCOUNT_USE_RECORDMCOUNT is enabled.
>
> This allows disabling just this option instead of disabling all FTRACE
> options.
>
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
>
> Or disabling FTRACE is the right thing to do if HAS_LTO_CLANG=y?
>
> Came from arch/Kconfig:
>
> config HAS_LTO_CLANG
>         def_bool y
>         depends on CC_IS_CLANG && LD_IS_LLD && AS_IS_LLVM
>         depends on $(success,$(NM) --help | head -n 1 | grep -qi llvm)
>         depends on $(success,$(AR) --help | head -n 1 | grep -qi llvm)
>         depends on ARCH_SUPPORTS_LTO_CLANG
>         depends on !FTRACE_MCOUNT_USE_RECORDMCOUNT  <======
>         depends on !KASAN || KASAN_HW_TAGS
>         depends on !GCOV_KERNEL
>
>
> ---
>  kernel/trace/Kconfig | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
> index a5eb5e7fd624..87d82d2b0b0b 100644
> --- a/kernel/trace/Kconfig
> +++ b/kernel/trace/Kconfig
> @@ -704,7 +704,8 @@ config FTRACE_MCOUNT_USE_OBJTOOL
>         depends on FTRACE_MCOUNT_RECORD
>
>  config FTRACE_MCOUNT_USE_RECORDMCOUNT
> -       def_bool y
> +       bool "Enable FTRACE_MCOUNT_USE_RECORDMCOUNT"
> +       default y
>         depends on !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY
>         depends on !FTRACE_MCOUNT_USE_CC
>         depends on !FTRACE_MCOUNT_USE_OBJTOOL
> --
> 2.30.2
>


-- 
Thanks,
~Nick Desaulniers

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Steven Rostedt]

On Fri, 11 Feb 2022 12:43:13 +1100
Alexey Kardashevskiy <aik@ozlabs.ru> wrote:

> For whatever reason LLVM does not allow LTO (Link Time Optimization) if
> FTRACE_MCOUNT_USE_RECORDMCOUNT is enabled.
> 
> This allows disabling just this option instead of disabling all FTRACE
> options.

What FTRACE options are you talking about?

> 
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
> 
> Or disabling FTRACE is the right thing to do if HAS_LTO_CLANG=y?
> 
> Came from arch/Kconfig:
> 
> config HAS_LTO_CLANG
>         def_bool y
>         depends on CC_IS_CLANG && LD_IS_LLD && AS_IS_LLVM
>         depends on $(success,$(NM) --help | head -n 1 | grep -qi llvm)
>         depends on $(success,$(AR) --help | head -n 1 | grep -qi llvm)
>         depends on ARCH_SUPPORTS_LTO_CLANG
>         depends on !FTRACE_MCOUNT_USE_RECORDMCOUNT  <======
>         depends on !KASAN || KASAN_HW_TAGS
>         depends on !GCOV_KERNEL
> 
> 
> ---
>  kernel/trace/Kconfig | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
> index a5eb5e7fd624..87d82d2b0b0b 100644
> --- a/kernel/trace/Kconfig
> +++ b/kernel/trace/Kconfig
> @@ -704,7 +704,8 @@ config FTRACE_MCOUNT_USE_OBJTOOL
>  	depends on FTRACE_MCOUNT_RECORD
>  
>  config FTRACE_MCOUNT_USE_RECORDMCOUNT
> -	def_bool y
> +	bool "Enable FTRACE_MCOUNT_USE_RECORDMCOUNT"
> +	default y

I don't think this does what you think it does.

This is not something that should be user selectable. What exactly are
you trying to accomplish here?

-- Steve


>  	depends on !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY
>  	depends on !FTRACE_MCOUNT_USE_CC
>  	depends on !FTRACE_MCOUNT_USE_OBJTOOL

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Sami Tolvanen]

On Fri, Feb 11, 2022 at 4:11 PM Nick Desaulniers
<ndesaulniers@google.com> wrote:
>
> + Sami, Kees
>
> Do you guys remember why we don't support
> FTRACE_MCOUNT_USE_RECORDMCOUNT with LTO?

It's because recordmcount cannot distinguish between calls and other
references to mcount/fentry, which is a problem with LTO where we have
to process the entire vmlinux.o and cannot ignore individual
translation units. Instead of resorting to workarounds, Peter was kind
enough to implement a smarter mcount pass in objtool, which we now use
on x86_64 with LTO.

Alexey, which architecture are you trying to fix here?

Sami

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Alexey Kardashevskiy]

On 2/12/22 13:47, Steven Rostedt wrote:
> On Fri, 11 Feb 2022 12:43:13 +1100
> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> 
>> For whatever reason LLVM does not allow LTO (Link Time Optimization) if
>> FTRACE_MCOUNT_USE_RECORDMCOUNT is enabled.
>>
>> This allows disabling just this option instead of disabling all FTRACE
>> options.
> 
> What FTRACE options are you talking about?

LTO_CLANG_THIN => HAS_LTO_CLANG => !FTRACE_MCOUNT_USE_RECORDMCOUNT =>

FTRACE [=y] && !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY [=n] && 
!FTRACE_MCOUNT_USE_CC [=n] && !FTRACE_MCOUNT_USE_OBJTOOL [=n] && 
FTRACE_MCOUNT_RECORD [=y]

A bunch.


> 
>>
>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>> ---
>>
>> Or disabling FTRACE is the right thing to do if HAS_LTO_CLANG=y?
>>
>> Came from arch/Kconfig:
>>
>> config HAS_LTO_CLANG
>>          def_bool y
>>          depends on CC_IS_CLANG && LD_IS_LLD && AS_IS_LLVM
>>          depends on $(success,$(NM) --help | head -n 1 | grep -qi llvm)
>>          depends on $(success,$(AR) --help | head -n 1 | grep -qi llvm)
>>          depends on ARCH_SUPPORTS_LTO_CLANG
>>          depends on !FTRACE_MCOUNT_USE_RECORDMCOUNT  <======
>>          depends on !KASAN || KASAN_HW_TAGS
>>          depends on !GCOV_KERNEL
>>
>>
>> ---
>>   kernel/trace/Kconfig | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
>> index a5eb5e7fd624..87d82d2b0b0b 100644
>> --- a/kernel/trace/Kconfig
>> +++ b/kernel/trace/Kconfig
>> @@ -704,7 +704,8 @@ config FTRACE_MCOUNT_USE_OBJTOOL
>>   	depends on FTRACE_MCOUNT_RECORD
>>   
>>   config FTRACE_MCOUNT_USE_RECORDMCOUNT
>> -	def_bool y
>> +	bool "Enable FTRACE_MCOUNT_USE_RECORDMCOUNT"
>> +	default y
> 
> I don't think this does what you think it does.

Sounds like it.

> This is not something that should be user selectable. What exactly are
> you trying to accomplish here?

I am trying to

diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
index b779603978e1..91c122224f83 100644
--- a/arch/powerpc/Kconfig
+++ b/arch/powerpc/Kconfig
@@ -153,6 +153,8 @@ config PPC
         select ARCH_WANT_IRQS_OFF_ACTIVATE_MM
         select ARCH_WANT_LD_ORPHAN_WARN
         select ARCH_WEAK_RELEASE_ACQUIRE
+       select ARCH_SUPPORTS_LTO_CLANG          if PPC64
+       select ARCH_SUPPORTS_LTO_CLANG_THIN     if PPC64


to get LTO working on powerpc64le with minimal change to 
ppc64le_defconfig which has all these FTRACE_xxx enabled.


> 
> -- Steve
> 
> 
>>   	depends on !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY
>>   	depends on !FTRACE_MCOUNT_USE_CC
>>   	depends on !FTRACE_MCOUNT_USE_OBJTOOL
> 

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Steven Rostedt]

On Wed, 16 Feb 2022 12:19:14 +1100
Alexey Kardashevskiy <aik@ozlabs.ru> wrote:

> On 2/12/22 13:47, Steven Rostedt wrote:
> > On Fri, 11 Feb 2022 12:43:13 +1100
> > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >   
> >> For whatever reason LLVM does not allow LTO (Link Time Optimization) if
> >> FTRACE_MCOUNT_USE_RECORDMCOUNT is enabled.
> >>
> >> This allows disabling just this option instead of disabling all FTRACE
> >> options.  
> > 
> > What FTRACE options are you talking about?  
> 
> LTO_CLANG_THIN => HAS_LTO_CLANG => !FTRACE_MCOUNT_USE_RECORDMCOUNT =>
> 
> FTRACE [=y] && !FTRACE_MCOUNT_USE_PATCHABLE_FUNCTION_ENTRY [=n] && 
> !FTRACE_MCOUNT_USE_CC [=n] && !FTRACE_MCOUNT_USE_OBJTOOL [=n] && 
> FTRACE_MCOUNT_RECORD [=y]

So basically, turn off FUNCTION_TRACER and STACK_TRACER.

> 
> A bunch.
> 
> 
> >   
> >>
> >> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> >> ---
> >>
> >> Or disabling FTRACE is the right thing to do if HAS_LTO_CLANG=y?
> >>
> >> Came from arch/Kconfig:
> >>
> >> config HAS_LTO_CLANG
> >>          def_bool y
> >>          depends on CC_IS_CLANG && LD_IS_LLD && AS_IS_LLVM
> >>          depends on $(success,$(NM) --help | head -n 1 | grep -qi llvm)
> >>          depends on $(success,$(AR) --help | head -n 1 | grep -qi llvm)
> >>          depends on ARCH_SUPPORTS_LTO_CLANG
> >>          depends on !FTRACE_MCOUNT_USE_RECORDMCOUNT  <======
> >>          depends on !KASAN || KASAN_HW_TAGS
> >>          depends on !GCOV_KERNEL
> >>
> >>
> >> ---
> >>   kernel/trace/Kconfig | 3 ++-
> >>   1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
> >> index a5eb5e7fd624..87d82d2b0b0b 100644
> >> --- a/kernel/trace/Kconfig
> >> +++ b/kernel/trace/Kconfig
> >> @@ -704,7 +704,8 @@ config FTRACE_MCOUNT_USE_OBJTOOL
> >>   	depends on FTRACE_MCOUNT_RECORD
> >>   
> >>   config FTRACE_MCOUNT_USE_RECORDMCOUNT
> >> -	def_bool y
> >> +	bool "Enable FTRACE_MCOUNT_USE_RECORDMCOUNT"
> >> +	default y  
> > 
> > I don't think this does what you think it does.  
> 
> Sounds like it.
> 
> > This is not something that should be user selectable. What exactly are
> > you trying to accomplish here?  
> 
> I am trying to
> 
> diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
> index b779603978e1..91c122224f83 100644
> --- a/arch/powerpc/Kconfig
> +++ b/arch/powerpc/Kconfig
> @@ -153,6 +153,8 @@ config PPC
>          select ARCH_WANT_IRQS_OFF_ACTIVATE_MM
>          select ARCH_WANT_LD_ORPHAN_WARN
>          select ARCH_WEAK_RELEASE_ACQUIRE
> +       select ARCH_SUPPORTS_LTO_CLANG          if PPC64
> +       select ARCH_SUPPORTS_LTO_CLANG_THIN     if PPC64
> 
> 
> to get LTO working on powerpc64le with minimal change to 
> ppc64le_defconfig which has all these FTRACE_xxx enabled.

Well, you can add a dependency against FUNCTION_TRACER. As that's what
turns on the RECORD_MCOUNT, as RECORD_MCOUNT is needed by it.

-- Steve

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Alexey Kardashevskiy]

On 2/15/22 03:10, Sami Tolvanen wrote:
> On Fri, Feb 11, 2022 at 4:11 PM Nick Desaulniers
> <ndesaulniers@google.com> wrote:
>>
>> + Sami, Kees
>>
>> Do you guys remember why we don't support
>> FTRACE_MCOUNT_USE_RECORDMCOUNT with LTO?
> 
> It's because recordmcount cannot distinguish between calls and other
> references to mcount/fentry, which is a problem with LTO where we have
> to process the entire vmlinux.o and cannot ignore individual
> translation units. Instead of resorting to workarounds, Peter was kind
> enough to implement a smarter mcount pass in objtool, which we now use
> on x86_64 with LTO.


So can x86_64 have both FTRACE_MCOUNT_USE_RECORDMCOUNT and LTO?

> Alexey, which architecture are you trying to fix here?


powerpc64le.

> 
> Sami

Re: [RFC PATCH kernel] trace: Make FTRACE_MCOUNT_USE_RECORDMCOUNT configurable

[Naveen N. Rao]

Alexey Kardashevskiy wrote:
> 
> 
> On 2/15/22 03:10, Sami Tolvanen wrote:
>> On Fri, Feb 11, 2022 at 4:11 PM Nick Desaulniers
>> <ndesaulniers@google.com> wrote:
>>>
>>> + Sami, Kees
>>>
>>> Do you guys remember why we don't support
>>> FTRACE_MCOUNT_USE_RECORDMCOUNT with LTO?
>> 
>> It's because recordmcount cannot distinguish between calls and other
>> references to mcount/fentry, which is a problem with LTO where we have
>> to process the entire vmlinux.o and cannot ignore individual
>> translation units. Instead of resorting to workarounds, Peter was kind
>> enough to implement a smarter mcount pass in objtool, which we now use
>> on x86_64 with LTO.

I had responded on a different thread about this:
http://lkml.kernel.org/r/1644579444.bal9kji6uc.naveen@linux.ibm.com

We are working on enabling 'objtool check --mcount' on powerpc.

> 
> 
> So can x86_64 have both FTRACE_MCOUNT_USE_RECORDMCOUNT and LTO?

No. x86_64 now supports HAVE_OBJTOOL_MCOUNT, which enables 
FTRACE_MCOUNT_USE_OBJTOOL. That allows LTO to be used.


- Naveen