binutils ld and new PT_GNU_PROPERTY segment

[Florian Weimer]

* Fangrui Song:

> Below is my understanding of these matters. Hope they will be useful for
> interested stakeholders (for example, AArch64 devs, though PT_GNU_PROPERTY is
> currently driven by x86) who don't follow the discussions so closely.
>
> 1. We need PT_GNU_PROPERTY.

>  Old linkers don't know the special processing on input .note.gnu.property sections.
>  The output .note.gnu.property does not take -z ibt/-z shstk/-z force-bti/-z pac-plt into account =>
>  invalid.
>  The produced PT_NOTE may contain multiple NT_GNU_PROPERTY_TYPE_0 => invalid [3]

In practice, we can recognize binaries produced by old linkers when the
object has been linked on a CET-enabled distribution because the
produced notes are always invalid.  The glibc dynamic loader already
checks for this and does not enable CET in this case.

ld -r involving exactly one CET-enabled object and one or more non-CET
objects is still problematic, of course, but it seems an unlikely
outcome.  The only way I can see this happening is with a CET-by-default
GCC (such as the one Ubuntu uses), but then you still would have to use
another linker (not /usr/bin/ld).  So even that seems like a fringe
issue to me.

In short, we looked at this situation, and still think that it's
supportable.

>  Also note that sh_addralign(.note.gnu.property)=8 on a 64-bit
>  platform, while
>  sh_addralign(.note.gnu.build-id)=sh_addralign(.note.ABI-tag)=...=4
>  (ancient mistake made by at least Linux/FreeBSD/NetBSD/...) GNU ld
>  before PR ld/23658 may create corrupted PT_NOTE.

Yes, we ran into some of these issues in our distribution and had to
rebuild a few objects.

>  For at least the above reasons, loaders are better not interpreting PT_NOTE.
>  glibc/sysdeps/x86/dl-prop.h is currently interpreting PT_NOTE => it
>  should be fixed.

I'm not sure we can do that for backwards compatibility reasons.  The
ABI has been out there for several years now.

>  Given point 1 and 3, this comment deserves a reconsideration:
>
>  > Binaries with .note.gnu.property section have been put into many
>  > OS releases.  We must support them.
>
> 2. .note.gnu.property behaves strangely, unlike a regular SHT_NOTE.
>   For a .note.gnu.property aware linker (newer GNU ld, newer lld),
>  .note.gnu.property input sections are dropped.
>
>  (We have .note.GNU-stack and .note.GNU-split-stack which both require special processing, but
>  they are SHT_PROGBITS.)
>
> 3. We need SHT_GNU_PROPERTY.
>  The output .note.gnu.property being SHT_NOTE causes linkers to place the section in both PT_NOTE
>  and PT_GNU_PROPERTY.
>  PT_NOTE, as explained by point 1 above, can cause trouble to old loaders.
>  Have we proved that "older linker-produced concatenated PT_NOTE cannot cause trouble to loaders interpreting PT_NOTE"?
>
>  SHT_GNU_PROPERTY does not contribute to PT_NOTE and will not cause any problem to old loaders
>  interpreting PT_NOTE.

Yes, I agree that it's desirable to add SHT_GNU_PROPERTY.

Thanks,
Florian