Re: [RFC PATCH 7/9] hurd: Generalize init-first.c to support x86_64

[Sergey Bugaev]

> I wish I had a better understanding of just what's going on in this
> file. Maybe a lot of the hacks can be rewritten in a nicer way. For
> instance, do we really need to hijack the return addresses and jump to
> init1 in this weird way, only to enable it to access argc/arg0? Since
> we know where they are on our stack (__builtin_frame_address (0) + 2
> or something like that), can't we just pass it a pointer?
>
> Let me actually try just that...

Eh, no, let's start even earlier than that. Please correct me if I
(inevitably) get things wrong.

_dl_init_first is Hurd-specific. It is called from the assembly code
in dl-machine.h, specifically there's a RTLD_START_SPECIAL_INIT macro
that's defined to call _dl_init_first on the Hurd, and to nothing
otherwise. This RTLD_START_SPECIAL_INIT is used ("invoked") in
i386/dl-machine.h, s390/s390-{32,64}/dl-machine.h, ia64/dl-machine.h,
and alpha/dl-machine.h (but notably not in x86_64/dl-machine.h). In
all cases, it's emphasized that "The special initializer gets called
with the stack just as the application's entry point will see it; it
can switch stacks if it moves these contents over."

But I conclude that:
- s390-gnu, ia64-gnu, and alpha-gnu ports are nonexistent, nor are
they realistically ever going to happen, so we can ignore them
completely
- the implementation does not seem to actually switch stacks (in fact,
I have removed the unused switch_stacks function in the last commit)
-- so the "gets called with the stack just as the application's entry
point will see it" property may not be important anymore?

The only thing it really needs, it seems, is a pointer to
argc/argv/envp & Hurd data block *somewhere*. It does not have to be
on the stack (though where else would it be), or immediately preceding
its call frame -- all that really matters is that there's a pointer.

So my thinking goes, why don't we just hook into _dl_start, which
already has this very pointer? And in fact, _dl_start calls
_dl_sysdep_start, for which there is already a Hurd version. Can't we
just call our logic from there, and not worry about the stack layout
and overwriting return addresses?

That would work for the SHARED case; we also need to do something for
the static case. In that case, we are invoked by static-start.S; do I
understand it right that the argc/argv/whatever is still located
on-stack even on x86_64 in this case, and not passed in registers
according to the calling convention? Then again, we should be able to
just use __builtin_frame_address (0) + 2 and avoid most of the hacks?

Please tell me if this makes any sense.

Sergey