Re: [PATCH] efi/libstub: refactor Makefile to not use lib-y syntax

From: Masahiro Yamada
Date: Wed Jun 03 2020 - 07:23:27 EST


On Wed, Jun 3, 2020 at 6:16 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
>
> On Wed, 3 Jun 2020 at 11:15, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> >
> > On Wed, Jun 3, 2020 at 6:02 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > >
> > > On Wed, 3 Jun 2020 at 10:59, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Jun 3, 2020 at 5:45 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Wed, 3 Jun 2020 at 10:36, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Wed, Jun 3, 2020 at 3:45 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Wed, 3 Jun 2020 at 07:34, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > > Documentation/kbuild/makefiles.rst says:
> > > > > > > >
> > > > > > > > Use of lib-y is normally restricted to `lib/` and `arch/*/lib`.
> > > > > > > >
> > > > > > > > I want to disallow lib-y outside of them.
> > > > > > > >
> > > > > > >
> > > > > > > Why?
> > > > > >
> > > > > >
> > > > > > Because I plan to remove lib-y entirely at some point.
> > > > > >
> > > > > > lib-y is not so useful to shrink the image size because:
> > > > > >
> > > > > > - An object in lib.a can be omitted only when no symbol
> > > > > > in that object is referenced. This rarely happens.
> > > > > >
> > > > > > - lib-y objects are often exported by nature
> > > > > > because lib-y is a collection of utility functions.
> > > > > > Even if no in-tree user, we always need to keep them
> > > > > > because EXPORT_SYMBOL() is the interface to modules.
> > > > > >
> > > > > >
> > > > > > When I worked on commit 7273ad2b08f8ac9563579d16a3cf528857b26f49,
> > > > > > I made some research.
> > > > > >
> > > > > > The benefit of lib-y is just 362 byte for x86_64_defconfig.
> > > > > > ( Before: 26578002, After: 26578364)
> > > > > >
> > > > > > My hope is lib-y will be replaced by dead-code elimination or
> > > > > > ultimately by LTO.
> > > > > >
> > > > > > drivers/firmware/efi/libstub/Makefile
> > > > > > is the only Makefile that breaks the rule:
> > > > > > "Use of lib-y is normally restricted to `lib/` and `arch/*/lib`"
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > >
> > > > > > > > Add a custom rule to build lib.a, which is linked to the decompressor
> > > > > > > > for ARCH=x86, ARCH=arm.
> > > > > > > >
> > > > > > > > For ARCH=arm64, use obj-y to link objects to vmlinux in the ordinary
> > > > > > > > way.
> > > > > > > >
> > > > > > >
> > > > > > > The code works perfectly fine as is, and I don't see what is
> > > > > > > fundamentally wrong with using static libraries outside of lib/ and
> > > > > > > arch/*/lib.
> > > > > >
> > > > > > The intended usage of lib-y is to hook lib.a
> > > > > > to scripts/vmlinux.sh via KBUILD_VMLINUX_LIBS.
> > > > > >
> > > > > > This Makefile is just what you found to work.
> > > > > >
> > > > > >
> > > > > > >
> > > > > > > Also, I would like this code to still be incorporated as a static
> > > > > > > library into arm64 as well, so that only pieces that are actually
> > > > > > > needed are incorporated into the final image.
> > > > > >
> > > > > > No.
> > > > > > It is not working like that because you set
> > > > > > lib.a to core-y.
> > > > > >
> > > > > > All objects in core-y are always linked to vmlinux.
> > > > > >
> > > > >
> > > > > The lib.a file is passed to the linker as a static library, so it will
> > > > > only grab what it needs.
> > > > >
> > > > > For instance, if you build arm64 from mainline today, the
> > > > > efi_relocate_kernel will not be in the final image, even though it is
> > > > > built as part of libstub
> > > >
> > > >
> > > > I built today's mainline kernel
> > > > (d6f9469a03d832dcd17041ed67774ffb5f3e73b3).
> > > >
> > > >
> > > > I see it in vmlinux.
> > > >
> > > >
> > > > $ make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- defconfig
> > > > $ make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- -j24
> > > > ...
> > > > $ aarch64-linux-gnu-nm -n vmlinux | grep efi_relocate_kernel
> > > > ffff8000114afb90 t __efistub_efi_relocate_kernel
> > > >
> > >
> > > That is strange. I tested this before, and it worked.
> > >
> > > Did anything change recently in the way the linker is invoked?
> >
> >
> > Nothing recently.
> >
> > This is obvious result because
> > drivers/firmware/efi/libstub/lib.a
> > is passed after ----whole-archive flag.
> >
> >
> > I guess the following commit is it,
> > but it is already 3 years ago.
> >
>
> Right.
>
> So that means there is no point at all in using static libraries, no?
> Or does --whole-archive still allow duplicate definitions like static
> libraries do usually?


--whole-archive literally links the whole of the archive.
If there are duplicate definitions, the linking fails
due to multiple definition.
So, --whole-archive is similar to linking .o files
in this regard.


There is a point for arm and x86 because
the decompressor link does not use --whole-archive.

As for arm64, there is no point,
but you can pass lib.a after --no-whole-archive
by the following patch.

diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile
index 650e1185c190..48a6afa774fc 100644
--- a/arch/arm64/Makefile
+++ b/arch/arm64/Makefile
@@ -145,7 +145,7 @@ export TEXT_OFFSET

core-y += arch/arm64/
libs-y := arch/arm64/lib/ $(libs-y)
-core-$(CONFIG_EFI_STUB) += $(objtree)/drivers/firmware/efi/libstub/lib.a
+libs-$(CONFIG_EFI_STUB) += $(objtree)/drivers/firmware/efi/libstub/lib.a

# Default target when executing plain make
boot := arch/arm64/boot



Then, _efistub_efi_low_alloc_above
and __efistub_efi_relocate_kernel will be dropped.


If you like the static library
do you want me to send v2?


--
Best Regards
Masahiro Yamada