Re: [PATCH 2/2] Discard .note.gnu.property sections in generic NOTES

From: Omar Sandoval
Date: Mon Sep 19 2022 - 13:27:09 EST

On Sat, Sep 17, 2022 at 06:31:20AM +0000, Christophe Leroy wrote:
>
>
> Le 16/09/2022 à 21:40, Omar Sandoval a écrit :
> > [Vous ne recevez pas souvent de courriers de osandov@xxxxxxxxxxx. D?couvrez pourquoi ceci est important ? https://aka.ms/LearnAboutSenderIdentification ]
> >
> > On Tue, Apr 28, 2020 at 06:21:05AM -0700, H.J. Lu wrote:
> >> With the command-line option, -mx86-used-note=yes, the x86 assembler
> >> in binutils 2.32 and above generates a program property note in a note
> >> section, .note.gnu.property, to encode used x86 ISAs and features. But
> >> kernel linker script only contains a single NOTE segment:
> >>
> >> PHDRS {
> >> text PT_LOAD FLAGS(5);
> >> data PT_LOAD FLAGS(6);
> >> percpu PT_LOAD FLAGS(6);
> >> init PT_LOAD FLAGS(7);
> >> note PT_NOTE FLAGS(0);
> >> }
> >> SECTIONS
> >> {
> >> ...
> >> .notes : AT(ADDR(.notes) - 0xffffffff80000000) { __start_notes = .; KEEP(*(.not
> >> e.*)) __stop_notes = .; } :text :note
> >> ...
> >> }
> >>
> >> The NOTE segment generated by kernel linker script is aligned to 4 bytes.
> >> But .note.gnu.property section must be aligned to 8 bytes on x86-64 and
> >> we get
> >>
> >> [hjl@gnu-skx-1 linux]$ readelf -n vmlinux
> >>
> >> Displaying notes found in: .notes
> >> Owner Data size Description
> >> Xen 0x00000006 Unknown note type: (0x00000006)
> >> description data: 6c 69 6e 75 78 00
> >> Xen 0x00000004 Unknown note type: (0x00000007)
> >> description data: 32 2e 36 00
> >> xen-3.0 0x00000005 Unknown note type: (0x006e6558)
> >> description data: 08 00 00 00 03
> >> readelf: Warning: note with invalid namesz and/or descsz found at offset 0x50
> >> readelf: Warning: type: 0xffffffff, namesize: 0x006e6558, descsize:
> >> 0x80000000, alignment: 8
> >> [hjl@gnu-skx-1 linux]$
> >>
> >> Since note.gnu.property section in kernel image is never used, this patch
> >> discards .note.gnu.property sections in kernel linker script by adding
> >>
> >> /DISCARD/ : {
> >> *(.note.gnu.property)
> >> }
> >>
> >> before kernel NOTE segment in generic NOTES.
> >>
> >> Signed-off-by: H.J. Lu <hjl.tools@xxxxxxxxx>
> >> Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
> >> ---
> >> include/asm-generic/vmlinux.lds.h | 7 +++++++
> >> 1 file changed, 7 insertions(+)
> >>
> >> diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
> >> index 71e387a5fe90..95cd678428f4 100644
> >> --- a/include/asm-generic/vmlinux.lds.h
> >> +++ b/include/asm-generic/vmlinux.lds.h
> >> @@ -833,7 +833,14 @@
> >> #define TRACEDATA
> >> #endif
> >>
> >> +/*
> >> + * Discard .note.gnu.property sections which are unused and have
> >> + * different alignment requirement from kernel note sections.
> >> + */
> >> #define NOTES \
> >> + /DISCARD/ : { \
> >> + *(.note.gnu.property) \
> >> + } \
> >> .notes : AT(ADDR(.notes) - LOAD_OFFSET) { \
> >> __start_notes = .; \
> >> KEEP(*(.note.*)) \
> >> --
> >> 2.25.4
> >>
> >
> > Hi, H.J.,
> >
> > I recently ran into this same .notes corruption when building kernels on
> > Arch Linux.
> >
> > What ended up happening to this patch? It doesn't appear to have been
> > merged, and I couldn't find any further discussion about it. I'm happy
> > to resend it for you if you need a hand.
>
> As far as I can see, ARM64 is doing something with that section, see
> arch/arm64/include/asm/assembler.h
>
> Instead of discarding that section, would it be enough to force
> alignment of .notes to 8 bytes ?
>
> Thanks
> Christophe

Unfortunately, "alignment requirement" here isn't just the starting
alignment of the .notes section; it also refers to internal padding in
the note metadata to keep things aligned. Changing this would break
anyone who parses /sys/kernel/notes (e.g., perf).

Here is a little more context around this mess:

The System V gABI [1] says that the note header and descriptor should be
aligned to 4 bytes for 32-bit files and 8 bytes for 64-bit files.
However, Linux never followed this, and 4-byte alignment is used for
both 32-bit and 64-bit files; see elf(5) [2].

The only exception as of 2022 is
.note.gnu.property/NT_GNU_PROPERTY_TYPE_0, which is defined to follow
the gABI alignment. There was a long thread discussing this back in 2018
with the subject "PT_NOTE alignment, NT_GNU_PROPERTY_TYPE_0, glibc and
gold" [3].

According to the gABI Linux Extensions [4], consumers are now supposed
to use the p_align of the PT_NOTE segment instead of assuming an
alignment.

There are a few issues with this for the kernel:

* The vmlinux linker script squishes together all of the notes sections
with different alignments without adjusting their internal padding,
but sets p_align to the maximum required alignment. This is what
confuses readelf and co: they expect 8-byte alignment, but most of the
note entries are only padded for 4-byte alignment.
* The vmlinux .notes section is exported as /sys/kernel/notes. This is
stable ABI and has always had 4-byte alignment; all existing parsers
assume this.
* /sys/kernel/notes doesn't currently have a way to specify an alternate
alignment anyways.

My suggestion would be to keep .note.gnu.property in its own section in
vmlinux, or create a new .notes8 section with 8-byte alignment, and
leave .notes and /sys/kernel/notes alone.

I'm not sure what exactly arch/arm64/include/asm/assembler.h is doing
with this file. Perhaps the author, Mark Brown, can clarify?

1: http://www.sco.com/developers/gabi/latest/ch5.pheader.html#note_section
2: https://man7.org/linux/man-pages/man5/elf.5.html#:~:text=Notes%20(Nhdr)
3: https://public-inbox.org/libc-alpha/13a92cb0-a993-f684-9a96-e02e4afb1bef@xxxxxxxxxx/
4: https://gitlab.com/x86-psABIs/Linux-ABI