Re: [RFC PATCH 03/23] Allow glibc to be compiled without EXEC_PAGESIZE

[Sergey Bugaev]

Hello,

On Wed, Apr 10, 2024 at 2:57 PM Florian Weimer <fweimer@redhat.com> wrote:
> * Sergey Bugaev:
>
> > We could define EXEC_PAGESIZE to some conservative value on
> > aarch64-gnu too, if it turns out that this little workaround is really
> > required. But it seems cleaner to make sure we don't need to, as
> > Roland's email suggests, and introducing a new port that doesn't have
> > a fixed page size (and doesn't come with an EXEC_PAGESIZE value
> > already defined in kernel headers) seems to be a good opportunity to
> > do that. That's my reasoning here.
>
> But the ELF image must be laid out with certain expectations regarding
> the maximum support page size.  Otherwise, something (kernel or dynamic
> linker) needs to perform copies or upgrade conflicting permissions
> within one page to a superset of all permissions.  I don't think we have
> code for that today, and we wouldn't necessarily want to implement that,
> I think.

Certainly -- and I wouldn't expect the kernel or the dynamic linker to
do anything about it other than reporting an error.

I think what you're saying is you consider EXEC_PAGESIZE to indeed
describe/define this required alignment of ELF segments (as the name
suggests). If I adopt that view, then yes, having EXEC_PAGESIZE makes
sense, and it makes some sense to use it as a conservative page size
value to use while the real value is not yet available (assuming there
is a real need for that).

The way I've been viewing it -- based on the fact that neither Linux's
nor glibc's (dynamic) nor BFD's nor LLVM's (static) linkers use it for
that purpose -- is that it's just some PAGE_SIZE-like definition
that's unrelated to binary loading (despite its name -- perhaps it has
been historically related to segment alignment in some old versions of
Linux?) that has been co-opted by glibc for pre-initializing
dl_pagesize, for dubious reasons. It also seems to be a Linux- (and
x86 Hurd) specific thing; I cannot find it in the BSDs.

Which one is it?

Sergey