On Thu, 2019-05-02 at 17:14 +0100, Dave Martin wrote:
> On Thu, May 02, 2019 at 08:47:06AM -0700, Yu-cheng Yu wrote:
> > On Thu, 2019-05-02 at 12:10 +0100, Dave Martin wrote:
> > > On Wed, May 01, 2019 at 02:12:17PM -0700, Yu-cheng Yu wrote:
> > > > An ELF file's .note.gnu.property indicates features the executable file
> > > > can support. For example, the property GNU_PROPERTY_X86_FEATURE_1_AND
> > > > indicates the file supports GNU_PROPERTY_X86_FEATURE_1_IBT and/or
> > > > GNU_PROPERTY_X86_FEATURE_1_SHSTK.
> >
> > [...]
> > > A couple of questions before I look in more detail:
> > >
> > > 1) Can we rely on PT_GNU_PROPERTY being present in the phdrs to describe
> > > the NT_GNU_PROPERTY_TYPE_0 note? If so, we can avoid trying to parse
> > > irrelevant PT_NOTE segments.
> >
> > Some older linkers can create multiples of NT_GNU_PROPERTY_TYPE_0. The code
> > scans all PT_NOTE segments to ensure there is only one
> > NT_GNU_PROPERTY_TYPE_0.
> > If there are multiples, then all are considered invalid.
>
> I'm concerned that in the arm64 case we would waste some effort by
> scanning multiple notes.
>
> Could we do something like iterating over the phdrs, and if we find
> exactly one PT_GNU_PROPERTY then use that, else fall back to scanning
> all PT_NOTEs?
That makes sense to me, but the concern is that we don't know the
PT_GNU_PROPERTY the only one. This probably needs to be discussed with more
people.
> > > 2) Are there standard types for things like the program property header?
> > > If not, can we add something in elf.h? We should try to coordinate with
> > > libc on that. Something like
> > >
> > > typedef __u32 Elf_Word;
> > >
> > > typedef struct {
> > > Elf_Word pr_type;
> > > Elf_Word pr_datasz;
> > > } Elf_Gnu_Prophdr;
> > >
> > > (i.e., just the header part from [1], with a more specific name -- which
> > > I just made up).
> >
> > Yes, I will fix that.
> >
> > [...]
> > > 3) It looks like we have to go and re-parse all the notes for every
> > > property requested by the arch code.
> >
> > As explained above, it is necessary to scan all PT_NOTE segments. But there
> > should be only one NT_GNU_PROPERTY_TYPE_0 in an ELF file. Once that is
> > found,
> > perhaps we can store it somewhere, or call into the arch code as you
> > mentioned
> > below. I will look into that.
>
> Just to get something working on arm64, I'm working on some hacks that
> move things around a bit -- I'll post when I have something.
>
> Did you have any view on my other point, below?
That should work. I will also make some changes for that.
>
> Cheers
> ---Dave
>
> > > For now there is only one property requested anyway, so this is probably
> > > not too bad. But could we flip things around so that we have some
> > > CONFIG_ARCH_WANTS_ELF_GNU_PROPERTY (say), and have the ELF core code
> > > call into the arch backend for each property found?
> > >
> > > The arch could provide some hook
> > >
> > > int arch_elf_has_gnu_property(const Elf_Gnu_Prophdr *prop,
> > > const void *data);
> > >
> > > to consume the properties as they are found.
> > >
> > > This would effectively replace the arch_setup_property() hook you
> > > currently have.
> > >
> > > Cheers
> > > ---Dave
> > >
> > > [1] https://github.com/hjl-tools/linux-abi/wiki/Linux-Extensions-to-gABI
