Re: [PATCH 5/5] add setting gs/fsbase

[Samuel Thibault]

Sergey Bugaev, le ven. 21 avril 2023 17:30:12 +0300, a ecrit:
> On Fri, Apr 21, 2023 at 12:24 PM Samuel Thibault
> <samuel.thibault@gnu.org> wrote:
> > No, they don't :) otherwise they wouldn't respect the ABI. The AMD64
> > ABI indeed defaults to /lib/ld64.so.1, and says that Linux uses
> > /lib64/ld-linux-x86-64.so.2, and that is cast in stone, we cannot change
> > it any more.
> >
> > /lib/ld-x86-64.so.1 should be completely fine.
> 
> I don't think I understand -- so the AMD64 ABI spec requires it to be
> /lib/ld64.so.1,

It does not require it to be that, it says it should be that. But it's a
very bad idea since it doesn't permit to distinguish different 64bit
archs.

> but it also mentions that Linux uses
> /lib64/ld-linux-x86-64.so.2,

Yes, because Linux chose to use that, to encode the OS name and the arch
in the path.

> and that means we should be using /lib/ld-x86-64.so.1?

Yes, because we do want to put the arch name in the path. We could want
to put "gnu-" in it too, it's just about time to decide. After we have
some settled distribution, we should really not change that.

> Moreover, how does the ELF spec requiring a specific file path for
> INTERP make any sense at all?

For interoperability between distributions.

> Surely the .1 is the soname means something, we should be able to bump
> it to .2 when we break ABI really bad?

Yes, but we'd rather avoid that if possible.

> What if the system is not a Unix and doesn't use /lib/ or .so for its
> libraries (but still uses ELF as a format)?

Then it's not the same system and thus it's free to use whatever
interpreter path it prefers. But it should really rather stick to it
once decided, for interoperability.

Samuel