Re: [PATCH v5 3/4] riscv: Add ifunc-compatible hwprobe function

[Florian Weimer <fweimer@redhat.com>]

* Evan Green:

> Palmer and I chatted a bit more about this idea. The evolution of it
> was to define something like AT_RISCV_HWPROBE_FUNC whose value is a
> pointer to the vDSO function. The __riscv_hwprobe_early() in
> libc_nonshared can then just get the auxval and call it, with similar
> semantics as we have here of not setting errno, etc.

I don't think it can access the auxiliary vector without relocations.

> Florian, one question about your other possible solution suggestion,
> reproduced below:
>
>> You could pass the function pointer to the IFUNC resolver
>> (which may require a marker symbol and GCC changes).
>
> I'm unsure what a marker symbol is, so I'm having trouble following
> this suggestion. I'd like to understand what that suggestion was,
> especially if you think it's better than what we're planning above.

POWER does this to signal that a binary needs data in the TCB for use by
IFUNC resolvers:

/* Newer LIBCs explicitly export this symbol to declare that they provide
   the AT_PLATFORM and AT_HWCAP/AT_HWCAP2 values in the TCB.  We emit a
   reference to this symbol whenever we expand a CPU builtin, so that
   we never link against an old LIBC.  */
const char *tcb_verification_symbol = "__parse_hwcap_and_convert_at_platform";

You are already passing an additional unused argument to IFUNC
resolvers, though:

static inline ElfW(Addr)
__attribute ((always_inline))
elf_ifunc_invoke (ElfW(Addr) addr)
{
  /* The second argument is a void pointer to preserve the extension
     fexibility.  */
  return ((ElfW(Addr) (*) (uint64_t, void *)) (addr))
         (GLRO(dl_hwcap), NULL);
}

So you could pass something there (like the address of __riscv_hwprobe),
and add a third argument (NULL for now) for future extensions.

This change does not even have ABI impact, applications just need to
remember to check the function pointer for NULL.  I haven't check if GCC
changes are needed before the compiler accepts the additional argument
in IFUNC resolvers.

Thanks,
Florian