Re: [PATCH] efi/libstub/x86: work around LLVM ELF quirk build regression

[Fangrui Song <maskray@google.com>]

On 2020-05-04, Nick Desaulniers wrote:
>On Mon, May 4, 2020 at 1:16 AM Ard Biesheuvel <ardb@kernel.org> wrote:
>>
>> When building the x86 EFI stub with Clang, the libstub Makefile rules
>> that manipulate the ELF object files may throw an error like:
>>
>>     STUBCPY drivers/firmware/efi/libstub/efi-stub-helper.stub.o
>>   strip: drivers/firmware/efi/libstub/efi-stub-helper.stub.o: Failed to find link section for section 10
>>   objcopy: drivers/firmware/efi/libstub/efi-stub-helper.stub.o: Failed to find link section for section 10
>>
>> This is the result of a LLVM 'feature' [0] where symbol references are
>> stored in a LLVM specific .llvm_addrsig section in a non-transparent way,
>> causing generic ELF tools such as strip or objcopy to choke on them.
>>
>> So drop the .llvm_addrsig section explicitly as well, to work around
>> this behavior.
>>
>> [0] https://sourceware.org/bugzilla/show_bug.cgi?id=23817
>
>This page also has info about the extension:
>https://llvm.org/docs/Extensions.html#sht-llvm-addrsig-section-address-significance-table
>Peter, do you know if it's possible to prevent the emission of this
>section from clang?  Scanning through the sources, it looks like it's
>set unconditionally during LTO, but I couldn't find where else?  Is
>this section required for more than LTO?

-faddrsig has been the default for many platforms since clang 7.
You can find the current default state on various platforms here:

https://github.com/llvm/llvm-project/blob/master/clang/lib/Driver/ToolChains/Clang.cpp#L6157

The table may have other benefits but currently the only use case is lld
--icf=safe , which is safer and better than gold's --icf=safe (which
relies on (hacky) relocation scanning).

The current way things work:
(1) clang always emits .llvm_addrsig
(2) if the linker is lld and --icf=safe is specified, .llvm_addrsig is read to perform safe ICF

makes it easy for people to try --icf=safe, otherwise people will need
to change CFLAGS as well to try the linker size optimization.

This section has nothing to do with LTO.

I am a bit surprised that certain strip can error "Failed to find link section for section 10"
They were expected to set sh_link to 0.

>We need a generic way to disable LLVM extensions when we're not using
>the LLVM binutils.  We have a couple cases where `-no-integrated-as`
>will prevent AsmStreamer from not using assembler extensions, but in
>this case it's the linker+objcopy+strip that don't work with the
>extensions.

objcopy -R .llvm_addrsig can be used when it is difficult to pass -fno-addrsig

(it is also used here
https://sourceware.org/git/?p=glibc.git;a=blobdiff;f=csu/Makefile;h=7460bcb0cf1e6cb296cf76d6e8ed9e43044f89f9;hp=f3498960f89e3b31f7cda6969e6eb3393a305241;hb=3628670a04f9a53586bd91c01588c4462b5e01d3;hpb=b9dab9c53496a8cd5bb18342eceff8a584c37a3e
)

If possible, use -fno-addrsig instead.

>>
>> Cc: Nick Desaulniers <ndesaulniers@google.com>
>> Reported-by: Arnd Bergmann <arnd@arndb.de>
>
>Do you have a link to the configs or report so we can repro?
>
>Also, scripts/get_maintainer.pl should recommend our list for patches
>mentioning clang or llvm, which is a wider audience that can help test
>and review.  I've been out sick much of the past week, so I appreciate
>the shared help with code review.  Of course if your intention was to
>be more discreet, I'm sorry I may have just messed that up.
>
>> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
>> ---
>>  drivers/firmware/efi/libstub/Makefile | 5 +++--
>>  1 file changed, 3 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/firmware/efi/libstub/Makefile b/drivers/firmware/efi/libstub/Makefile
>> index 8d246b51bd49..4d137615a666 100644
>> --- a/drivers/firmware/efi/libstub/Makefile
>> +++ b/drivers/firmware/efi/libstub/Makefile
>> @@ -66,7 +66,8 @@ CFLAGS_arm64-stub.o           := -DTEXT_OFFSET=$(TEXT_OFFSET)
>>  # .data section of the compressed kernel to ensure initialization. Rename the
>>  # .bss section here so it's easy to pick out in the linker script.
>>  #
>> -STUBCOPY_FLAGS-$(CONFIG_X86)   += --rename-section .bss=.bss.efistub,load,alloc
>> +STUBCOPY_FLAGS-$(CONFIG_X86)   += -R .llvm_addrsig \
>> +                                  --rename-section .bss=.bss.efistub,load,alloc
>
>Do we only observe this for x86, not ARM?
>
>>  STUBCOPY_RELOC-$(CONFIG_X86_32)        := R_386_32
>>  STUBCOPY_RELOC-$(CONFIG_X86_64)        := R_X86_64_64
>>
>> @@ -111,7 +112,7 @@ $(obj)/%.stub.o: $(obj)/%.o FORCE
>>  #
>>  quiet_cmd_stubcopy = STUBCPY $@
>>        cmd_stubcopy =                                                   \
>> -       $(STRIP) --strip-debug -o $@ $<;                                \
>> +       $(STRIP) --strip-debug -R .llvm_addrsig -o $@ $<;               \
>>         if $(OBJDUMP) -r $@ | grep $(STUBCOPY_RELOC-y); then            \
>>                 echo "$@: absolute symbol references not allowed in the EFI stub" >&2; \
>>                 /bin/false;                                             \
>> --
>> 2.17.1
>>
>
>In the absence of a linker script where we can use `DISCARD` rules,
>this looks like the best approach, though I'm still curious about ARM.
>Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
>-- 
>Thanks,
>~Nick Desaulniers