Creating a default-visibility, weak, versioned alias for a function declared hidden

Creating a default-visibility, weak, versioned alias for a function declared hidden

[Zack Weinberg]

I need a little help wrangling the symbol version macros.

Suppose an existing function which is used internally to libc and also
a public API.  For concreteness, let's say it has this definition:

---
extern int frob (int fd);  // in frob/frob.h
extern typeof (frob) __frob attribute_hidden;  // in include/frob.h

// unix/finux/frob.c
int
__frob (int fd)
{
  // Older kernels don't do this check, POSIX requires it.
  if (!__isatty (fd))
    {
      __set_errno (ENOTTY);
      return -1;
    }
  return INLINE_SYSCALL_CALL (frob, fd);
}
weak_alias (__frob, frob);
---

This produces a frob.os with symbol table entries like this (readelf -Ws):

    22: 0000000000000000   136 FUNC    GLOBAL HIDDEN     1 __frob
    23: 0000000000000000   136 FUNC    WEAK   DEFAULT    1 frob

which is what we want.  But now suppose that, for reasons, some
architecture needs to override the default version of the public
symbol to be GLIBC_2.1 (there is a compat symbol at GLIBC_2.0).
Furthermore, on this architecture, `default_symbol_version (frob,
frob, GLIBC_2.1)` is rejected by the assembler. After some scratching
my head over libc-symbols.h I came up with this:

weak_alias (__frob, __frob_w);
default_symbol_version (__frob_w, frob, GLIBC_2.1);

But this produces a _hidden_ alias and versioned symbol:

    22: 0000000000000000   136 FUNC    GLOBAL HIDDEN     1 __frob
    23: 0000000000000000   136 FUNC    WEAK   HIDDEN     1 __frob_2
    24: 0000000000000000   136 FUNC    WEAK   HIDDEN     1 frob@@GLIBC_2.1

and external uses of 'frob' fail to link.

1) Why does `weak_alias` produce a default-visibility symbol in the
first case, but a hidden-visibility symbol in the second case?

2) How do I get the public symbol to be default visibility, weak, and versioned?

Thanks,
zw

Re: Creating a default-visibility, weak, versioned alias for a function declared hidden

[Carlos O'Donell]

On 7/29/19 9:14 AM, Zack Weinberg wrote:
> I need a little help wrangling the symbol version macros.
> 
> Suppose an existing function which is used internally to libc and also
> a public API.  For concreteness, let's say it has this definition:
> 
> ---
> extern int frob (int fd);  // in frob/frob.h
> extern typeof (frob) __frob attribute_hidden;  // in include/frob.h
> 
> // unix/finux/frob.c
> int
> __frob (int fd)
> {
>    // Older kernels don't do this check, POSIX requires it.
>    if (!__isatty (fd))
>      {
>        __set_errno (ENOTTY);
>        return -1;
>      }
>    return INLINE_SYSCALL_CALL (frob, fd);
> }
> weak_alias (__frob, frob);
> ---
> 
> This produces a frob.os with symbol table entries like this (readelf -Ws):
> 
>      22: 0000000000000000   136 FUNC    GLOBAL HIDDEN     1 __frob
>      23: 0000000000000000   136 FUNC    WEAK   DEFAULT    1 frob
> 
> which is what we want.  But now suppose that, for reasons, some
> architecture needs to override the default version of the public
> symbol to be GLIBC_2.1 (there is a compat symbol at GLIBC_2.0).
> Furthermore, on this architecture, `default_symbol_version (frob,
> frob, GLIBC_2.1)` is rejected by the assembler. After some scratching
> my head over libc-symbols.h I came up with this:
> 
> weak_alias (__frob, __frob_w);
> default_symbol_version (__frob_w, frob, GLIBC_2.1);
> 
> But this produces a _hidden_ alias and versioned symbol:
> 
>      22: 0000000000000000   136 FUNC    GLOBAL HIDDEN     1 __frob
>      23: 0000000000000000   136 FUNC    WEAK   HIDDEN     1 __frob_2
>      24: 0000000000000000   136 FUNC    WEAK   HIDDEN     1 frob@@GLIBC_2.1
> 
> and external uses of 'frob' fail to link.
> 
> 1) Why does `weak_alias` produce a default-visibility symbol in the
> first case, but a hidden-visibility symbol in the second case?

No idea. This surprises me too. The macros aren't the most robust in terms
of mixing and matching.

> 2) How do I get the public symbol to be default visibility, weak, and versioned?

This has to be possible because we have PLT bypass for versioned symbols.
I would just look up the existing examples to see what we did?

-- 
Cheers,
Carlos.