Re: [PATCH 1/2] arch/x86: Use -fno-asynchronous-unwind-tables to suppress .eh_frame sections

From: Fangrui Song
Date: Mon Feb 24 2020 - 16:12:16 EST


On 2020-02-24, Arvind Sankar wrote:
On Mon, Feb 24, 2020 at 12:33:49PM -0800, Nick Desaulniers wrote:
On Sun, Feb 23, 2020 at 11:37 AM Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote:
>
> While discussing a patch to discard .eh_frame from the compressed
> vmlinux using the linker script, Fangrui Song pointed out [1] that these
> sections shouldn't exist in the first place because arch/x86/Makefile
> uses -fno-asynchronous-unwind-tables.

Another benefit is that -fno-asynchronous-unwind-tables may help
reduce the size of .text!
https://stackoverflow.com/a/26302715/1027966

Hm I don't see any change in .text size.
> diff --git a/drivers/firmware/efi/libstub/Makefile b/drivers/firmware/efi/libstub/Makefile
> index 98a81576213d..a1140c4ee478 100644
> --- a/drivers/firmware/efi/libstub/Makefile
> +++ b/drivers/firmware/efi/libstub/Makefile
> @@ -12,7 +12,8 @@ cflags-$(CONFIG_X86) += -m$(BITS) -D__KERNEL__ -O2 \
> -mno-mmx -mno-sse -fshort-wchar \
> -Wno-pointer-sign \
> $(call cc-disable-warning, address-of-packed-member) \
> - $(call cc-disable-warning, gnu)
> + $(call cc-disable-warning, gnu) \
> + -fno-asynchronous-unwind-tables

I think we want to add this flag a little lower, line 27 has:

KBUILD_CFLAGS := $(cflags-y) -DDISABLE_BRANCH_PROFILING \

so the `cflags-y` variable you modify in this hunk will only set
-fno-asynchronous-unwind-tables for CONFIG_X86, which I don't think is
intentional. Though when I run

It is intentional -- the other case is that we're building for ARM,
which only filters out the regular KBUILD_CFLAGS, so adding the flag for
it should not be necessary. The cflags for ARM are constructed by
manipulating KBUILD_CFLAGS. Besides it may or may not want unwind
tables. 32-bit ARM appears to have an option to enable -funwind-tables.

clang (as of today) has not implemented the
-funwind-tables/-fasynchronous-unwind-tables distinction as GCC does..
(probably because not many people care..)


$ llvm-readelf -S drivers/firmware/efi/libstub/lib.a | grep eh_frame

after doing an x86_64 defconfig, I don't get any hits. Do you observe
.eh_frame sections on any of these objects in this dir? (I'm fine
adding it to be safe, but I'm curious why I'm not seeing any
.eh_frame)


You mean before this patch, right? I see hits on every .o file in there
(compiling with gcc 9.2.0).

>
> # arm64 uses the full KBUILD_CFLAGS so it's necessary to explicitly
> # disable the stackleak plugin
> --
> 2.24.1
>