Re: [PATCH] kbuild: Use --strip-unneeded with INSTALL_MOD_STRIP

[Charlie Jenkins <charlie@rivosinc.com>]

On Sat, Feb 01, 2025 at 12:10:02AM +0900, Masahiro Yamada wrote:
> On Fri, Jan 31, 2025 at 3:54 PM Charlie Jenkins <charlie@rivosinc.com> wrote:
> >
> > On Thu, Jan 30, 2025 at 08:52:45PM -0700, Nathan Chancellor wrote:
> > > On Wed, Jan 22, 2025 at 07:17:26PM -0800, Charlie Jenkins wrote:
> > > > On riscv, kernel modules end up with a significant number of local
> > > > symbols. This becomes apparent when compiling modules with debug symbols
> > > > enabled. Using amdgpu.ko as an example of a large module, on riscv the
> > > > size is 754MB (no stripping), 53MB (--strip-debug), and 21MB
> > > > (--strip-unneeded). ON x86, amdgpu.ko is 482MB (no stripping), 21MB
> > > > (--strip-debug), and 20MB (--strip-unneeded).
> > > >
> > > > Use --strip-unneeded instead of --strip-debug to strip modules so
> > > > decrease the size of the resulting modules. This is particularly
> > > > relevant for riscv, but also marginally aids other architectures.
> > > >
> > > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
> > >
> > > Is there any sort of regression risk with this patch? If so, another
> > > option may be to give another level to INSTALL_MOD_STRIP like 2 so that
> > > INSTALL_MOD_STRIP=1 continues to behave as before but people can easily
> > > opt into this option. No strong opinion because I am not sure but was
> > > not sure if it was considered.
> >
> > I do not believe this would cause regressions. The description on gnu
> > strip is:
> >
> > "Remove all symbols that are not needed for relocation processing in
> > addition to debugging symbols and sections stripped by --strip-debug."
> >
> > The description on llvm-strip is:
> >
> > "Remove from the output all local or undefined symbols that are not
> > required by relocations. Also remove all debug sections."
> >
> > gnu strip --strip-unneeded strips slightly more aggressively but it does
> > not appear this causes any issues.
> >
> > >
> > > Regardless:
> > >
> > > Reviewed-by: Nathan Chancellor <nathan@kernel.org>
> >
> > Thanks!
> >
> 
> 
> It is true --strip-unneeded drops a lot of compiler-generated symbols, but
> it also drops real symbols that originate in the source code.
> 
> So, this would give user-visible changes for kallsyms at least.

Adding INSTALL_MOD_STRIP="--strip-unneeded" would be sufficient for
riscv. However, this has the downside that riscv will require different
flags than other architectures to get reasonably sized modules. 

I believe these symbols are only useful for debugging, is there a
usecase for them to be available when the user has modules compiled with
INSTALL_MOD_STRIP=1?

- Charlie

> 
> 
> $ riscv64-linux-gnu-nm  -n
> /tmp/strip-unneeded/lib/modules/6.13.0-09760-g69e858e0b8b2/kernel/drivers/gpu/drm/amd/amdgpu/amdgpu.ko
>  > /tmp/symbol-with-strip-unneeded
> $ riscv64-linux-gnu-nm  -n
> /tmp/strip-debug/lib/modules/6.13.0-09760-g69e858e0b8b2/kernel/drivers/gpu/drm/amd/amdgpu/amdgpu.ko
>  > /tmp/symbol-with-strip-debug
> 
> $  diff -u /tmp/symbol-with-strip-debug  /tmp/symbol-with-strip-unneeded
>  [ snip ]
>  00000000001676cc t uvd_v6_0_ring_test_ring
>  0000000000167802 t uvd_v6_0_ring_emit_pipeline_sync
>  0000000000167a02 t uvd_v6_0_ring_emit_fence
> -0000000000167b58 r CSWTCH.2
> -0000000000167b68 r abm_settings
> -0000000000167b80 r abm_config
> -0000000000167b90 r min_reduction_table_v_2_2
> -0000000000167ba0 r max_reduction_table_v_2_2
> -0000000000167bb0 r min_reduction_table
> -0000000000167bc0 r max_reduction_table
> -0000000000167bd0 r custom_backlight_curve0
>  0000000000167c38 r abm_settings_config2
>  0000000000167c70 r abm_settings_config1
>  0000000000167ca8 r abm_settings_config0
> 
> 
> 
> 
> --
> Best Regards
> Masahiro Yamada