Re: [PATCH v2] tracing: Generate undef symbols allowlist for simple_ring_buffer

[Vincent Donnefort <vdonnefort@google.com>]

On Fri, Mar 13, 2026 at 09:37:24AM -0700, Nathan Chancellor wrote:
> On Fri, Mar 13, 2026 at 10:58:29AM +0000, Vincent Donnefort wrote:
> > Compiler and tooling-generated symbols are difficult to maintain
> > across all supported architectures. Make the allowlist more robust by
> > replacing the harcoded list with a mechanism that automatically detects
> > these symbols.
> > 
> > This mechanism generates a C function designed to trigger common
> > compiler-inserted symbols.
> > 
> > Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
> > 
> > ---
> > 
> > Changes in v2:
> > 
> >   - Use filechk (Nathan)
> >   - Removed deprecated extra-y (Nathan)
> >   - Added simple_ring_buffer in allowlist (Nathan)
> >   - Added memcpy() to generate more symbols (Nathan)
> >   - Added __sancov 
> > 
> > diff --git a/kernel/trace/Makefile b/kernel/trace/Makefile
> > index beb15936829d..96627a909ecc 100644
> > --- a/kernel/trace/Makefile
> > +++ b/kernel/trace/Makefile
> > @@ -136,17 +136,42 @@ obj-$(CONFIG_TRACE_REMOTE_TEST) += remote_test.o
> >  # simple_ring_buffer is used by the pKVM hypervisor which does not have access
> >  # to all kernel symbols. Fail the build if forbidden symbols are found.
> >  #
> > -UNDEFINED_ALLOWLIST := memset alt_cb_patch_nops __x86 __ubsan __asan __kasan __gcov __aeabi_unwind
> > -UNDEFINED_ALLOWLIST += __stack_chk_fail stackleak_track_stack __ref_stack __sanitizer llvm_gcda llvm_gcov
> > -UNDEFINED_ALLOWLIST += .TOC\. __clear_pages_unrolled __memmove copy_page warn_slowpath_fmt
> > -UNDEFINED_ALLOWLIST += ftrace_likely_update __hwasan_load __hwasan_store __hwasan_tag_memory
> > -UNDEFINED_ALLOWLIST += warn_bogus_irq_restore __stack_chk_guard
> > -UNDEFINED_ALLOWLIST := $(addprefix -e , $(UNDEFINED_ALLOWLIST))
> > +# undefsyms_base generates a set of compiler and tooling-generated symbols that can
> > +# safely be ignored for simple_ring_buffer.
> > +#
> > +filechk_undefsyms_base = \
> > +	echo '$(pound)include <linux/atomic.h>'; \
> > +	echo '$(pound)include <linux/string.h>'; \
> > +	echo '$(pound)include <asm/page.h>'; \
> > +	echo 'static char page[PAGE_SIZE] __aligned(PAGE_SIZE);'; \
> > +	echo 'void undefsyms_base(void *p, int n);'; \
> > +	echo 'void undefsyms_base(void *p, int n) {'; \
> > +	echo '	char buffer[256] = { 0 };'; \
> > +	echo '	u32 u = 0;'; \
> > +	echo '	memset((char * volatile)page, 8, PAGE_SIZE);'; \
> > +	echo '	memset((char * volatile)buffer, 8, sizeof(buffer));'; \
> > +	echo '	memcpy((void * volatile)p, buffer, sizeof(buffer));'; \
> > +	echo '	cmpxchg((u32 * volatile)&u, 0, 8);'; \
> > +	echo '	WARN_ON(n == 0xdeadbeef);'; \
> > +	echo '}'
> > +
> > +$(obj)/undefsyms_base.c: FORCE
> > +	$(call filechk,undefsyms_base)
> > +
> > +clean-files += undefsyms_base.c
> > +
> > +$(obj)/undefsyms_base.o: $(obj)/undefsyms_base.c
> > +
> > +targets += undefsyms_base.o
> > +
> > +UNDEFINED_ALLOWLIST = __asan __gcov __kasan __kcsan __hwasan __sancov __sanitizer __tsan __ubsan __x86_indirect_thunk \
> > +		      simple_ring_buffer \
> > +		      $(shell $(NM) -u $(obj)/undefsyms_base.o 2>/dev/null | awk '{print $$2}')
> >  
> >  quiet_cmd_check_undefined = NM      $<
> > -      cmd_check_undefined = test -z "`$(NM) -u $< | grep -v $(UNDEFINED_ALLOWLIST)`"
> > +      cmd_check_undefined = test -z "`$(NM) -u $< | grep -v $(addprefix -e , $(UNDEFINED_ALLOWLIST))`"
> >  
> > -$(obj)/%.o.checked: $(obj)/%.o FORCE
> > +$(obj)/%.o.checked: $(obj)/%.o $(obj)/undefsyms_base.o FORCE
> >  	$(call if_changed,check_undefined)
> >  
> >  always-$(CONFIG_SIMPLE_RING_BUFFER) += simple_ring_buffer.o.checked
> > 
> > base-commit: 33f2e266515717c4b2df585dadefa0525557726c
> > -- 
> > 2.53.0.851.ga537e3e6e9-goog
> > 
> 
> Thanks! This is almost perfect for my tests, one final thing that I
> noticed as a result of my full overnight builds. For ARCH=riscv (and
> some other architectures from a quick grep), there is some logic in
> their include/asm/string.h files to avoid FORTIFY_SOURCE when KASAN is
> enabled for the entire build but not enabled for the particular file. As
> undefsyms_base.o is not linked into vmlinux or modules, it does not
> automatically have KASAN enabled.
> 
>   $ cat allmod.config
>   CONFIG_GCOV_KERNEL=n
>   CONFIG_LTO_CLANG_THIN=y
>   CONFIG_WERROR=n
> 
>   $ make -skj"$(nproc)" ARCH=riscv KCONFIG_ALLCONFIG=1 LLVM=1 mrproper allmodconfig kernel/trace/
>   Unexpected symbols in kernel/trace/simple_ring_buffer.o:
>                    U __fortify_panic
>                    U __write_overflow_field
>   ...
> 
> This cures that for me.
> 
> diff --git a/kernel/trace/Makefile b/kernel/trace/Makefile
> index 260382f62dbf..55af887a90e2 100644
> --- a/kernel/trace/Makefile
> +++ b/kernel/trace/Makefile
> @@ -164,6 +164,11 @@ $(obj)/undefsyms_base.o: $(obj)/undefsyms_base.c
>  
>  targets += undefsyms_base.o
>  
> +# ensure KASAN is enabled to avoid logic that may disable FORTIFY_SOURCE when
> +# KASAN is not enabled. undefsyms_base.o does not automatically get KASAN flags
> +# because it is not linked into vmlinux.
> +KASAN_SANITIZE_undefsyms_base.o := y
> +
>  UNDEFINED_ALLOWLIST = __asan __gcov __kasan __kcsan __hwasan __sancov __sanitizer __tsan __ubsan __x86_indirect_thunk \
>  		      simple_ring_buffer \
>  		      $(shell $(NM) -u $(obj)/undefsyms_base.o 2>/dev/null | awk '{print $$2}')
> --
> 
> With that addressed:
> 
> Reviewed-by: Nathan Chancellor <nathan@kernel.org>
> Tested-by: Nathan Chancellor <nathan@kernel.org>

I've just sent a v3 with all that. I have tested locally with allmodconfig and
many architectures with both clang and gcc.

Thanks a lot for your help!

--
Vincent

> 
> Cheers,
> Nathan