Re: [PATCH v2 0/7] linkage: better symbol aliasing

[Ard Biesheuvel]

On Tue, 25 Jan 2022 at 16:46, Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> On Tue, Jan 25, 2022 at 04:28:11PM +0100, Ard Biesheuvel wrote:
> > On Tue, 25 Jan 2022 at 12:32, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > >
> > > This series aims to make symbol aliasing simpler and more consistent.
> > > The basic idea is to replace SYM_FUNC_START_ALIAS(alias) and
> > > SYM_FUNC_END_ALIAS(alias) with a new SYM_FUNC_ALIAS(alias, name), so
> > > that e.g.
> > >
> > >     SYM_FUNC_START(func)
> > >     SYM_FUNC_START_ALIAS(alias1)
> > >     SYM_FUNC_START_ALIAS(alias2)
> > >         ... asm insns ...
> > >     SYM_FUNC_END(func)
> > >     SYM_FUNC_END_ALIAS(alias1)
> > >     SYM_FUNC_END_ALIAS(alias2)
> > >     EXPORT_SYMBOL(alias1)
> > >     EXPORT_SYMBOL(alias2)
> > >
> > > ... can become:
> > >
> > >     SYM_FUNC_START(name)
> > >         ... asm insns ...
> > >     SYM_FUNC_END(name)
> > >
> > >     SYM_FUNC_ALIAS(alias1, func)
> > >     EXPORT_SYMBOL(alias1)
> > >
> > >     SYM_FUNC_ALIAS(alias2, func)
> > >     EXPORT_SYMBOL(alias2)
> > >
> > > This avoids repetition and hopefully make it easier to ensure
> > > consistency (e.g. so each function has a single canonical name and
> > > associated metadata).
> > >
> >
> > I take it this affects the sizes of the alias ELF symbols? Does that matter?
>
> The alias should be given the same size as the original symbol, unless I've
> made an error. If you look at patch 3:
>
> * In SYM_FUNC_START(name), via SYM_ENTRY_AT(name, ...), we create a local label
>   for the start of the function: .L____sym_entry__##name
>
> * In SYM_FUNC_END(name), via SYM_END_AT(name, ...), we create a local label for
>   the end of the function: .L____sym_end__##name
>
> * In SYM_FUNC_ALIAS*(alias,name), we define the start and end of the alias as
>   the start and end of the original symbol using those local labels, e.g.
>
>   | #define SYM_FUNC_ALIAS(alias, name)                                     \
>   |         SYM_START_AT(alias, .L____sym_entry__##name, SYM_L_GLOBAL)      \
>   |         SYM_END_AT(alias, .L____sym_end__##name, SYM_T_FUNC)
>
> Note that:
>
> * SYM_FUNC_START() uses SYM_START(), which uses SYM_ENTRY_AT()
> * SYM_FUNC_END() uses SYM_END(), which uses SYM_END_AT()
>
> ... so the definition of tha alias is ultimately structurally identical to the
> definition of the canoncial name, at least for now.
>

Ah right, apologies for not looking more carefully - I assumed the
changed placement implied that the aliases had zero size.

And ultimately, I don't think there is an obviously correct answer
anyway, it's just the [apparently non-existent] change in behavior I
was curious about.