Re: [PATCH bpf v2 1/2] compiler_types.h: Define __retain for __attribute__((__retain__))

[Tony Ambardar <tony.ambardar@gmail.com>]

On Fri, Jun 14, 2024 at 11:47:19AM -0700, Yonghong Song wrote:
> 
> On 6/10/24 3:56 PM, Tony Ambardar wrote:
> > On Tue, Jun 04, 2024 at 10:55:39PM -0700, Yonghong Song wrote:
> > > On 6/3/24 10:23 PM, Tony Ambardar wrote:
> > > > Some code includes the __used macro to prevent functions and data from
> > > > being optimized out. This macro implements __attribute__((__used__)), which
> > > > operates at the compiler and IR-level, and so still allows a linker to
> > > > remove objects intended to be kept.
> > > > 
> > > > Compilers supporting __attribute__((__retain__)) can address this gap by
> > > > setting the flag SHF_GNU_RETAIN on the section of a function/variable,
> > > > indicating to the linker the object should be retained. This attribute is
> > > > available since gcc 11, clang 13, and binutils 2.36.
> > > > 
> > > > Provide a __retain macro implementing __attribute__((__retain__)), whose
> > > > first user will be the '__bpf_kfunc' tag.
> > > > 
> > > > Link: https://lore.kernel.org/bpf/ZlmGoT9KiYLZd91S@krava/T/
> > > > Cc: stable@vger.kernel.org # v6.6+
> > > > Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com>
> > > > ---
> > > >    include/linux/compiler_types.h | 23 +++++++++++++++++++++++
> > > >    1 file changed, 23 insertions(+)
> > > > 
> > > > diff --git a/include/linux/compiler_types.h b/include/linux/compiler_types.h
> > > > index 93600de3800b..f14c275950b5 100644
> > > > --- a/include/linux/compiler_types.h
> > > > +++ b/include/linux/compiler_types.h
> > > > @@ -143,6 +143,29 @@ static inline void __chk_io_ptr(const volatile void __iomem *ptr) { }
> > > >    # define __preserve_most
> > > >    #endif
> > > > +/*
> > > > + * Annotating a function/variable with __retain tells the compiler to place
> > > > + * the object in its own section and set the flag SHF_GNU_RETAIN. This flag
> > > > + * instructs the linker to retain the object during garbage-cleanup or LTO
> > > > + * phases.
> > > > + *
> > > > + * Note that the __used macro is also used to prevent functions or data
> > > > + * being optimized out, but operates at the compiler/IR-level and may still
> > > > + * allow unintended removal of objects during linking.
> > > > + *
> > > > + * Optional: only supported since gcc >= 11, clang >= 13
> > > > + *
> > > > + *   gcc: https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#index-retain-function-attribute
> > > > + * clang: https://clang.llvm.org/docs/AttributeReference.html#retain
> > > > + */
> > > > +#if __has_attribute(__retain__) && \
> > > > +	(defined(CONFIG_LD_DEAD_CODE_DATA_ELIMINATION) || \
> > > > +	 defined(CONFIG_LTO_CLANG))
> > > Could you explain why CONFIG_LTO_CLANG is added here?
> > > IIUC, the __used macro permits garbage collection at section
> > > level, so CLANG_LTO_CLANG without
> > > CONFIG_LD_DEAD_CODE_DATA_ELIMINATION
> > > shuold not change final section dynamics, right?
> > Hi Yonghong,
> > 
> > I included the conditional guard to ensure consistent behaviour between
> > __retain and other features forcing split sections. In particular, the same
> > guard is used in vmlinux.lds.h to merge split sections where needed. For
> > example, using __retain in llvm builds without CONFIG_LTO was failing CI
> > tests on kernel-patches/bpf because the kernel didn't boot properly. And in
> > further testing, the kernel had no issues loading BPF kfunc modules with
> > such split sections, so I left the module (partial) linking scripts alone.
> 
> I tried with both bpf and bpf-next tree and I cannot make CONFIG_HAVE_LD_DEAD_CODE_DATA_ELIMINATION=y
> in .config file. The following are all occurances in Kconfig:

My understanding is one doesn't directly set HAVE_LD_DEAD_CODE_...; it's a
per-arch capability flag which guards setting LD_DEAD_CODE_DATA_ELIMINATION
but only targets "small systems" (i.e. embedded), so no surprise x86 isn't
in the arch list below.

> 
> $ egrep -r HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> arch/mips/Kconfig:      select HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> arch/powerpc/Kconfig:   select HAVE_LD_DEAD_CODE_DATA_ELIMINATION if HAVE_OBJTOOL_MCOUNT && (!ARCH_USING_PATCHABLE_FUNCTION_ENTRY || (!CC_IS_GCC || GCC_VERSION >= 110100))
> arch/riscv/Kconfig:     select HAVE_LD_DEAD_CODE_DATA_ELIMINATION if !LD_IS_LLD
> init/Kconfig:config HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> init/Kconfig:   depends on HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> 
> Are there some pending patches to enable HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> for x86?

I doubt it given the target arches above, but curious what's the need for
x86 support? Only x86_32? My patches were motivated seeing resolve_btfids
and pahole errors for a couple years on MIPS routers. I don't recall seeing
the same for x86 builds, so my testing focussed more on preserving x86
builds rather than adding/testing the arch flag for x86.

> 
> I could foce CONFIG_HAVE_LD_DEAD_CODE_DATA_ELIMINATION=y with the following hack:
> diff --git a/init/Kconfig b/init/Kconfig
> index 72404c1f2157..adf8718e2f5b 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -1402,7 +1402,7 @@ config CC_OPTIMIZE_FOR_SIZE
>  endchoice
>  config HAVE_LD_DEAD_CODE_DATA_ELIMINATION
> -       bool
> +       def_bool y
>         help
>           This requires that the arch annotates or otherwise protects
>           its external entry points from being discarded. Linker scripts
> 
> But with the above, I cannot boot the kernel.

OK, interesting exercise. Setting HAVE_LD_DEAD_CODE_DATA_ELIMINATION
shouldn't change anything itself so I suppose you are also setting
LD_DEAD_CODE_DATA_ELIMINATION? From previous testing on kernel-patches/CI,
first guess would be vmlinux linker script doing section merges unaware of
some x86 quirk. Or x86-specific linker script unhappy with split sections.

> 
> 
> Did I miss anything?
> 
> > 
> > Maybe I misunderstand you question re: __used?
> > 
> > Thanks,
> > Tony
> > > > +# define __retain			__attribute__((__retain__))
> > > > +#else
> > > > +# define __retain
> > > > +#endif
> > > > +
> > > >    /* Compiler specific macros. */
> > > >    #ifdef __clang__
> > > >    #include <linux/compiler-clang.h>