[PATCH] arm64: remove special treatment for the link order of head.o

[Andrii Nakryiko]

On Fri, Mar 24, 2023 at 4:34 PM Alexei Starovoitov
<alexei.starovoitov at gmail.com> wrote:
>
> On Fri, Mar 24, 2023 at 4:39 AM Ard Biesheuvel <ardb at kernel.org> wrote:
> >
> > (cc BTF list and maintainer)
> >
> > On Thu, 23 Mar 2023 at 22:12, Aurelien Jarno <aurelien at aurel32.net> wrote:
> > >
> > > Hi,
> > >
> > > On 2023-03-22 15:51, Ard Biesheuvel wrote:
> > > > On Tue, 21 Mar 2023 at 23:26, Aurelien Jarno <aurelien at aurel32.net> wrote:
> > > > >
> > > > > Hi,
> > > > >
> > > > > On 2022-10-13 08:35, Masahiro Yamada wrote:
> > > > > > In the previous discussion (see the Link tag), Ard pointed out that
> > > > > > arm/arm64/kernel/head.o does not need any special treatment - the only
> > > > > > piece that must appear right at the start of the binary image is the
> > > > > > image header which is emitted into .head.text.
> > > > > >
> > > > > > The linker script does the right thing to do. The build system does
> > > > > > not need to manipulate the link order of head.o.
> > > > > >
> > > > > > Link: https://lore.kernel.org/lkml/CAMj1kXH77Ja8bSsq2Qj8Ck9iSZKw=1F8Uy-uAWGVDm4-CG=EuA@mail.gmail.com/
> > > > > > Suggested-by: Ard Biesheuvel <ardb at kernel.org>
> > > > > > Signed-off-by: Masahiro Yamada <masahiroy at kernel.org>
> > > > > > ---
> > > > > >
> > > > > >  scripts/head-object-list.txt | 1 -
> > > > > >  1 file changed, 1 deletion(-)
> > > > > >
> > > > > > diff --git a/scripts/head-object-list.txt b/scripts/head-object-list.txt
> > > > > > index b16326a92c45..f226e45e3b7b 100644
> > > > > > --- a/scripts/head-object-list.txt
> > > > > > +++ b/scripts/head-object-list.txt
> > > > > > @@ -15,7 +15,6 @@ arch/alpha/kernel/head.o
> > > > > >  arch/arc/kernel/head.o
> > > > > >  arch/arm/kernel/head-nommu.o
> > > > > >  arch/arm/kernel/head.o
> > > > > > -arch/arm64/kernel/head.o
> > > > > >  arch/csky/kernel/head.o
> > > > > >  arch/hexagon/kernel/head.o
> > > > > >  arch/ia64/kernel/head.o
> > > > >
> > > > > This patch causes a significant increase of the arch/arm64/boot/Image
> > > > > size. For instance the generic arm64 Debian kernel went from 31 to 39 MB
> > > > > after this patch has been applied to the 6.1 stable tree.
> > > > >
> > > > > In turn this causes issues with some bootloaders, for instance U-Boot on
> > > > > a Raspberry Pi limits the kernel size to 36 MB.
> > > > >
> > > >
> > > > I cannot reproduce this with mainline
> > > >
> > > > With the patch
> > > >
> > > > $ size vmlinux
> > > >    text    data     bss     dec     hex filename
> > > > 24567309 14752630 621680 39941619 26175f3 vmlinux
> > > >
> > > > With the patch reverted
> > > >
> > > > $ size vmlinux
> > > >    text    data     bss     dec     hex filename
> > > > 24567309 14752694 621680 39941683 2617633 vmlinux
> > >
> > > I have tried with the current mainline, this is what I get, using GCC 12.2.0
> > > and binutils 2.40:
> > >
> > >    text    data     bss     dec     hex filename
> > > 32531655        8192996  621968 41346619        276e63b vmlinux.orig
> > > 25170610        8192996  621968 33985574        2069426 vmlinux.revert
> > >
> > > > It would help to compare the resulting vmlinux ELF images from both
> > > > builds to see where the extra space is being allocated
> > >
> > > At a first glance, it seems the extra space is allocated in the BTF
> > > section. I have uploaded the resulting files as well as the config file
> > > I used there:
> > > https://temp.aurel32.net/linux-arm64-size-head.o.tar.gz
> > >
> >
> > Indeed. So we go from
> >
> >   [15] .BTF              PROGBITS         ffff8000091d1ff4  011e1ff4
> >        00000000005093d6  0000000000000000   A       0     0     1
> >
> > to
> >
> >   [15] .BTF              PROGBITS         ffff8000091d1ff4  011e1ff4
> >        0000000000c0e5eb  0000000000000000   A       0     0     1
> >
> > i.e, from 5 MiB to 12+ MiB of BTF metadata.
> >
> > To me, it is not clear at all how one would be related to the other,
> > so it will leave it to the Kbuild and BTF experts to chew on this one.
>
> That's a huge increase.
> It's not just that commit responsible, but the whole series ?
> https://lore.kernel.org/lkml/20220906061313.1445810-1-masahiroy@kernel.org/
> I'm guessing "Link vmlinux and modules in parallel" is related.
> I'm not sure what "parallel link" means. Running 'ar' in parallel?
> I cannot read makefile syntax, so no idea.
>
> Jiri, Andrii, Alan, please take a look.

So it seems to come from the difference in return type for mm_struct's
get_unmapped_area callback:


struct mm_struct {
        struct {
                struct maple_tree mm_mt;
#ifdef CONFIG_MMU
                unsigned long (*get_unmapped_area) (struct file *filp,
                                unsigned long addr, unsigned long len,
                                unsigned long pgoff, unsigned long flags);
#endif


It seems that sometimes we have "unsigned long" as return type, but
sometimes it's just "void". I haven't debugged why this is happening.
But cc'ing Eduard, just in case it's related to the "unspecified type"
tracking in pahole, which he recently fixed. There could be some other
related bug lurking around.