On Tue, Feb 13, 2024 at 09:55:07AM +0900, Masahiro Yamada wrote:
> On Fri, Feb 9, 2024 at 5:21 AM Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
> >
> > Commit 90ceddcb4950 ("bpf: Support llvm-objcopy for vmlinux BTF")
> > changed the ELF type of .btf.vmlinux.bin.o from ET_EXEC to ET_REL via
> > dd, which works fine for little endian platforms:
> >
> > 00000000 7f 45 4c 46 02 01 01 00 00 00 00 00 00 00 00 00 |.ELF............|
> > -00000010 03 00 b7 00 01 00 00 00 00 00 00 80 00 80 ff ff |................|
>
> I am afraid this dump is confusing.
>
> The byte stream "03 00" is ET_DYN, as specified in ELF:
>
> Name Value
> ------------------
> ET_REL 1
> ET_EXEC 2
> ET_DYN 3
>
> It disagrees with your commit message "from ET_EXEC to ET_REL"
>
> The dump for the old ELF was "02 00", wasn't it?
No, I have not manually edited or changed these diffs from hexdumping
the .o files. The little endian one was from arm64 and the big endian
one was from s390. Perhaps this is some difference between the
toolchains? I don't recall which one I used in this case, pretty sure it
was GNU though. I can just remove "from ET_EXEC" from the commit message
if that would help make it less confusing.
> > +00000010 01 00 b7 00 01 00 00 00 00 00 00 80 00 80 ff ff |................|
> >
> > However, for big endian platforms, it changes the wrong byte, resulting
> > in an invalid ELF file type, which ld.lld rejects:
>
> Fangrui pointed out this is true for inutils >= 2.35
Not for this particular error, which occurs because e_type is not a
valid value. If it was true for binutils, we would have seen this issue
sooner.
> >
> > 00000000 7f 45 4c 46 02 02 01 00 00 00 00 00 00 00 00 00 |.ELF............|
> > -00000010 00 03 00 16 00 00 00 01 00 00 00 00 00 10 00 00 |................|
> > +00000010 01 03 00 16 00 00 00 01 00 00 00 00 00 10 00 00 |................|
>
> - 00 02
> + 01 02
See above.
> >
> > Type: <unknown>: 103
> >
> > ld.lld: error: .btf.vmlinux.bin.o: unknown file type
> >
> > Fix this by using a different seek value for dd when targeting big
> > endian, so that the correct byte gets changed and everything works
> > correctly for all linkers.
> >
> > 00000000 7f 45 4c 46 02 02 01 00 00 00 00 00 00 00 00 00 |.ELF............|
> > -00000010 00 03 00 16 00 00 00 01 00 00 00 00 00 10 00 00 |................|
>
> Ditto.
See above.
> > +00000010 00 01 00 16 00 00 00 01 00 00 00 00 00 10 00 00 |................|
> >
> > Type: REL (Relocatable file)
> >
> > Cc: stable@xxxxxxxxxxxxxxx
> > Fixes: 90ceddcb4950 ("bpf: Support llvm-objcopy for vmlinux BTF")
> > Link: https://github.com/llvm/llvm-project/pull/75643
> > Signed-off-by: Nathan Chancellor <nathan@xxxxxxxxxx>
> > ---
> > scripts/link-vmlinux.sh | 11 +++++++++--
> > 1 file changed, 9 insertions(+), 2 deletions(-)
> >
> > diff --git a/scripts/link-vmlinux.sh b/scripts/link-vmlinux.sh
> > index a432b171be82..8a9f48b3cb32 100755
> > --- a/scripts/link-vmlinux.sh
> > +++ b/scripts/link-vmlinux.sh
> > @@ -135,8 +135,15 @@ gen_btf()
> > ${OBJCOPY} --only-section=.BTF --set-section-flags .BTF=alloc,readonly \
> > --strip-all ${1} ${2} 2>/dev/null
> > # Change e_type to ET_REL so that it can be used to link final vmlinux.
> > - # Unlike GNU ld, lld does not allow an ET_EXEC input.
> > - printf '\1' | dd of=${2} conv=notrunc bs=1 seek=16 status=none
> > + # Unlike GNU ld, lld does not allow an ET_EXEC input. Make sure the correct
> > + # byte gets changed with big endian platforms, otherwise e_type may be an
> > + # invalid value.
> > + if is_enabled CONFIG_CPU_BIG_ENDIAN; then
> > + seek=17
> > + else
> > + seek=16
> > + fi
> > + printf '\1' | dd of=${2} conv=notrunc bs=1 seek=${seek} status=none
> > }
> >
> > # Create ${2} .S file with all symbols from the ${1} object file
>
> Do you want to send v2 to update the commit description?
>
I don't think a v2 is necessary for the commit description...
> The current code will work, but another approach might be to
> update both byte 16 and byte 17 because e_type is a 16-bit field.
>
> It works without relying on the MSB of the previous e_type being zero.
> The comment does not need updating because the intention is obvious
> from the code.
>
> if is_enabled CONFIG_CPU_BIG_ENDIAN; then
> et_rel='\0\1'
> else
> et_rel='\1\0'
> fi
>
> printf "${et_rel}" | dd of=${2} conv=notrunc bs=1 seek=16 status=none
but I do like this suggested change because I was thinking that updating
the single bit could be fragile at some point. I'll send a v2 with that
and a slightly updated commit message shortly. Because it is
substantially different from v1, I won't carry forward all the tags I
received but I hope people will take a look at v2 and provide them
again.
Thanks a lot for taking a look!
Nathan
