Don't create sframes during build

[RFC] Don't create sframes during build

Posted by Sebastian Andrzej Siewior 4 weeks ago

Hi,

gcc in Debian, starting with 15.2.0-2, 14.3.0-6 enables sframe
generation. Unless options like -ffreestanding are passed. Since this
isn't done, there are a few warnings during compile
| crypto/chacha.o: warning: objtool: .sframe+0x30: data relocation to !ENDBR: chacha_stream_xor+0x0
| crypto/chacha.o: warning: objtool: .sframe+0x94: data relocation to !ENDBR: crypto_xchacha_crypt+0x0

followed by warnings at the end

|   AR      vmlinux.a
|   LD      vmlinux.o
| vmlinux.o: warning: objtool: .sframe+0x15c: data relocation to !ENDBR: repair_env_string+0x0
| vmlinux.o: warning: objtool: .sframe+0x1c0: data relocation to !ENDBR: run_init_process+0x0
| vmlinux.o: warning: objtool: .sframe+0x1d4: data relocation to !ENDBR: try_to_run_init_process+0x0
| vmlinux.o: warning: objtool: .sframe+0x1e8: data relocation to !ENDBR: rcu_read_unlock+0x0
…
| vmlinux.o: warning: objtool: .sframe+0x12765c: data relocation to !ENDBR: get_eff_addr_reg+0x0
| vmlinux.o: warning: objtool: .sframe+0x1276ac: data relocation to !ENDBR: get_seg_base_limit+0x0
|   OBJCOPY modules.builtin.modinfo

followed by a boom
|   LD      .tmp_vmlinux1
| ld: error: unplaced orphan section `.sframe' from `vmlinux.o'

We could drop the sframe during the final link but this does not get rid
of the objtool warnings so we would have to ignore them. But we don't
need it. So what about the following:

diff --git a/Makefile b/Makefile
--- a/Makefile
+++ b/Makefile
@@ -886,6 +886,8 @@ ifdef CONFIG_CC_IS_GCC
 KBUILD_CFLAGS	+= $(call cc-option,--param=allow-store-data-races=0)
 KBUILD_CFLAGS	+= $(call cc-option,-fno-allow-store-data-races)
 endif
+# No sframe generation for kernel if enabled by default
+KBUILD_CFLAGS	+= $(call cc-option,-Xassembler --gsframe=no)
 
 ifdef CONFIG_READABLE_ASM
 # Disable optimizations that make assembler listings hard to read.
-- 
2.51.0

Re: [RFC] Don't create sframes during build

Posted by Matthias Klose 4 weeks ago

[ CCing binutils@sourceware.org ]

On 9/4/25 15:18, Sebastian Andrzej Siewior wrote:
> Hi,
> 
> gcc in Debian, starting with 15.2.0-2, 14.3.0-6 enables sframe
> generation. Unless options like -ffreestanding are passed. Since this
> isn't done, there are a few warnings during compile

If there are other options when sframe shouldn't be enabled, please tell.

Gentoo chose another approach, enabling sframe unconditionally in gas, 
unless disabled by --gsframe=no.

> | crypto/chacha.o: warning: objtool: .sframe+0x30: data relocation to !ENDBR: chacha_stream_xor+0x0
> | crypto/chacha.o: warning: objtool: .sframe+0x94: data relocation to !ENDBR: crypto_xchacha_crypt+0x0
> 
> followed by warnings at the end
> 
> |   AR      vmlinux.a
> |   LD      vmlinux.o
> | vmlinux.o: warning: objtool: .sframe+0x15c: data relocation to !ENDBR: repair_env_string+0x0
> | vmlinux.o: warning: objtool: .sframe+0x1c0: data relocation to !ENDBR: run_init_process+0x0
> | vmlinux.o: warning: objtool: .sframe+0x1d4: data relocation to !ENDBR: try_to_run_init_process+0x0
> | vmlinux.o: warning: objtool: .sframe+0x1e8: data relocation to !ENDBR: rcu_read_unlock+0x0
> …
> | vmlinux.o: warning: objtool: .sframe+0x12765c: data relocation to !ENDBR: get_eff_addr_reg+0x0
> | vmlinux.o: warning: objtool: .sframe+0x1276ac: data relocation to !ENDBR: get_seg_base_limit+0x0
> |   OBJCOPY modules.builtin.modinfo
> 
> followed by a boom
> |   LD      .tmp_vmlinux1
> | ld: error: unplaced orphan section `.sframe' from `vmlinux.o'
> 
> We could drop the sframe during the final link but this does not get rid
> of the objtool warnings so we would have to ignore them. But we don't
> need it. So what about the following:
> 
> diff --git a/Makefile b/Makefile
> --- a/Makefile
> +++ b/Makefile
> @@ -886,6 +886,8 @@ ifdef CONFIG_CC_IS_GCC
>   KBUILD_CFLAGS	+= $(call cc-option,--param=allow-store-data-races=0)
>   KBUILD_CFLAGS	+= $(call cc-option,-fno-allow-store-data-races)
>   endif
> +# No sframe generation for kernel if enabled by default
> +KBUILD_CFLAGS	+= $(call cc-option,-Xassembler --gsframe=no)
>   
>   ifdef CONFIG_READABLE_ASM
>   # Disable optimizations that make assembler listings hard to read.
This is what I chose for package builds that need disablement of sframe.

Re: [RFC] Don't create sframes during build

Posted by Sebastian Andrzej Siewior 4 weeks ago

On 2025-09-04 16:02:42 [+0200], Matthias Klose wrote:
> [ CCing binutils@sourceware.org ]
> 
> On 9/4/25 15:18, Sebastian Andrzej Siewior wrote:
> > Hi,
> > 
> > gcc in Debian, starting with 15.2.0-2, 14.3.0-6 enables sframe
> > generation. Unless options like -ffreestanding are passed. Since this
> > isn't done, there are a few warnings during compile
> 
> If there are other options when sframe shouldn't be enabled, please tell.

No, I think this is okay.

…
> > We could drop the sframe during the final link but this does not get rid
> > of the objtool warnings so we would have to ignore them. But we don't
> > need it. So what about the following:
> > 
> > diff --git a/Makefile b/Makefile
> > --- a/Makefile
> > +++ b/Makefile
> > @@ -886,6 +886,8 @@ ifdef CONFIG_CC_IS_GCC
> >   KBUILD_CFLAGS	+= $(call cc-option,--param=allow-store-data-races=0)
> >   KBUILD_CFLAGS	+= $(call cc-option,-fno-allow-store-data-races)
> >   endif
> > +# No sframe generation for kernel if enabled by default
> > +KBUILD_CFLAGS	+= $(call cc-option,-Xassembler --gsframe=no)
> >   ifdef CONFIG_READABLE_ASM
> >   # Disable optimizations that make assembler listings hard to read.
> This is what I chose for package builds that need disablement of sframe.

I think this would work for now. Longterm we would have to allow sframe
creation and keep section if an architecture decides to use it for its
backtracing. While orc seems fine on x86, there are arm64 patches to use
for as a stack unwinder.

Sebastian

Re: [RFC] Don't create sframes during build

Posted by Josh Poimboeuf 4 weeks ago

On Thu, Sep 04, 2025 at 06:34:04PM +0200, Sebastian Andrzej Siewior wrote:
> On 2025-09-04 16:02:42 [+0200], Matthias Klose wrote:
> > [ CCing binutils@sourceware.org ]
> > 
> > On 9/4/25 15:18, Sebastian Andrzej Siewior wrote:
> > > Hi,
> > > 
> > > gcc in Debian, starting with 15.2.0-2, 14.3.0-6 enables sframe
> > > generation. Unless options like -ffreestanding are passed. Since this
> > > isn't done, there are a few warnings during compile
> > 
> > If there are other options when sframe shouldn't be enabled, please tell.
> 
> No, I think this is okay.
> 
> …
> > > We could drop the sframe during the final link but this does not get rid
> > > of the objtool warnings so we would have to ignore them. But we don't
> > > need it. So what about the following:
> > > 
> > > diff --git a/Makefile b/Makefile
> > > --- a/Makefile
> > > +++ b/Makefile
> > > @@ -886,6 +886,8 @@ ifdef CONFIG_CC_IS_GCC
> > >   KBUILD_CFLAGS	+= $(call cc-option,--param=allow-store-data-races=0)
> > >   KBUILD_CFLAGS	+= $(call cc-option,-fno-allow-store-data-races)
> > >   endif
> > > +# No sframe generation for kernel if enabled by default
> > > +KBUILD_CFLAGS	+= $(call cc-option,-Xassembler --gsframe=no)
> > >   ifdef CONFIG_READABLE_ASM
> > >   # Disable optimizations that make assembler listings hard to read.
> > This is what I chose for package builds that need disablement of sframe.
> 
> I think this would work for now. Longterm we would have to allow sframe
> creation and keep section if an architecture decides to use it for its
> backtracing. While orc seems fine on x86, there are arm64 patches to use
> for as a stack unwinder.

This is probably fine, but... how does this interact with other kernel
makefiles enabling sframe?  For example, x86 will soon have a patch to
enable sframe generation for vdso.  And as you mentioned, arm64 will
enable it kernel-wide.

Removing the objtool !ENDBR warnings would be trivial (and is a good
thing to do regardless).

-- 
Josh

Re: [RFC] Don't create sframes during build

Posted by Matthias Klose 4 weeks ago

On 9/4/25 16:02, Matthias Klose wrote:
> [ CCing binutils@sourceware.org ]
> 
> On 9/4/25 15:18, Sebastian Andrzej Siewior wrote:
>> Hi,
>>
>> gcc in Debian, starting with 15.2.0-2, 14.3.0-6 enables sframe
>> generation. Unless options like -ffreestanding are passed. Since this
>> isn't done, there are a few warnings during compile
> 
> If there are other options when sframe shouldn't be enabled, please tell.
> 
> Gentoo chose another approach, enabling sframe unconditionally in gas, 
> unless disabled by --gsframe=no.
> 
>> | crypto/chacha.o: warning: objtool: .sframe+0x30: data relocation 
>> to !ENDBR: chacha_stream_xor+0x0
>> | crypto/chacha.o: warning: objtool: .sframe+0x94: data relocation 
>> to !ENDBR: crypto_xchacha_crypt+0x0
>>
>> followed by warnings at the end
>>
>> |   AR      vmlinux.a
>> |   LD      vmlinux.o
>> | vmlinux.o: warning: objtool: .sframe+0x15c: data relocation to ! 
>> ENDBR: repair_env_string+0x0
>> | vmlinux.o: warning: objtool: .sframe+0x1c0: data relocation to ! 
>> ENDBR: run_init_process+0x0
>> | vmlinux.o: warning: objtool: .sframe+0x1d4: data relocation to ! 
>> ENDBR: try_to_run_init_process+0x0
>> | vmlinux.o: warning: objtool: .sframe+0x1e8: data relocation to ! 
>> ENDBR: rcu_read_unlock+0x0
>> …
>> | vmlinux.o: warning: objtool: .sframe+0x12765c: data relocation to ! 
>> ENDBR: get_eff_addr_reg+0x0
>> | vmlinux.o: warning: objtool: .sframe+0x1276ac: data relocation to ! 
>> ENDBR: get_seg_base_limit+0x0
>> |   OBJCOPY modules.builtin.modinfo
>>
>> followed by a boom
>> |   LD      .tmp_vmlinux1
>> | ld: error: unplaced orphan section `.sframe' from `vmlinux.o'
>>
>> We could drop the sframe during the final link but this does not get rid
>> of the objtool warnings so we would have to ignore them. But we don't
>> need it. So what about the following:
>>
>> diff --git a/Makefile b/Makefile
>> --- a/Makefile
>> +++ b/Makefile
>> @@ -886,6 +886,8 @@ ifdef CONFIG_CC_IS_GCC
>>   KBUILD_CFLAGS    += $(call cc-option,--param=allow-store-data-races=0)
>>   KBUILD_CFLAGS    += $(call cc-option,-fno-allow-store-data-races)
>>   endif
>> +# No sframe generation for kernel if enabled by default
>> +KBUILD_CFLAGS    += $(call cc-option,-Xassembler --gsframe=no)
>>   ifdef CONFIG_READABLE_ASM
>>   # Disable optimizations that make assembler listings hard to read.
> This is what I chose for package builds that need disablement of sframe.


Fyi, I will disable that again next week, until sframe v3 is ready.