Re: [PATCH 1/4] x86: Initialize CPU info via IFUNC relocation [BZ 26203]

["H.J. Lu" <hjl.tools@gmail.com>]

On Mon, Sep 28, 2020 at 7:05 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu:
>
> >> > diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> >> > index 217c21c34f..7a325ab70e 100644
> >> > --- a/sysdeps/x86/cacheinfo.c
> >> > +++ b/sysdeps/x86/cacheinfo.c
> >>
> >> > +  assert (cpu_features->basic.kind != arch_kind_unknown);
> >>
> >> Why doesn't this assert fire occasionally?  How do you ensure that
> >
> > See
> >
> > https://sourceware.org/bugzilla/show_bug.cgi?id=26203
> >
> > It only happens in dlopen from a static executable.
>
> Sorry, I don't understand how this answers my question.
>
> Do you mean that for the non-static case, initialization has already
> happened.

Yes, it is initialized by DL_PLATFORM_INIT in a dynamic executable.

> >> relocation processing is correctly ordered?
> >
> > cpu_features is also initialized by IFUNC relocation in ld.so which
> > is relocated before libc.so.
>
> Is that really true in all cases?  Even if libc.so is preloaded?

Not if init_cacheinfo is a constructor function.

> (Static dlopen probably ignores LD_PRELOAD.)

Correct.

> Maybe put this information as a comment next to the assert?

Will do.

> But since cacheinfo.os is linked into libc.so, I don't really think the
> assert is correct.

When init_cacheinfo is called, cpu_features must be initialized.

> >> > diff --git a/sysdeps/x86/dl-get-cpu-features.c b/sysdeps/x86/dl-get-cpu-features.c
> >> > index 5f9e46b0c6..da4697b895 100644
> >> > --- a/sysdeps/x86/dl-get-cpu-features.c
> >> > +++ b/sysdeps/x86/dl-get-cpu-features.c
> >> > @@ -1,4 +1,4 @@
> >> > -/* This file is part of the GNU C Library.
> >> > +/* Initialize CPU feature data via IFUNC relocation.
> >> >     Copyright (C) 2015-2020 Free Software Foundation, Inc.
> >> >
> >> >     The GNU C Library is free software; you can redistribute it and/or
> >> > @@ -18,6 +18,29 @@
> >> >
> >> >  #include <ldsodefs.h>
> >> >
> >> > +#ifdef SHARED
> >> > +# include <cpu-features.c>
> >> > +
> >> > +/* NB: Normally, DL_PLATFORM_INIT calls init_cpu_features to initialize
> >> > +   CPU features.  But when loading ld.so inside of static executable,
> >> > +   DL_PLATFORM_INIT isn't called.  Call init_cpu_features by initializing
> >> > +   a dummy function pointer via IFUNC relocation for ld.so.  */
> >> > +extern void __x86_cpu_features (void) attribute_hidden;
> >> > +const void (*__x86_cpu_features_p) (void) attribute_hidden
> >> > +  = __x86_cpu_features;
> >> > +
> >> > +void
> >> > +_dl_x86_init_cpu_features (void)
> >> > +{
> >> > +  struct cpu_features *cpu_features = __get_cpu_features ();
> >> > +  if (cpu_features->basic.kind == arch_kind_unknown)
> >> > +    init_cpu_features (cpu_features);
> >> > +}
> >> > +
> >> > +__ifunc (__x86_cpu_features, __x86_cpu_features, NULL, void,
> >> > +      _dl_x86_init_cpu_features);
> >> > +#endif
> >> > +
> >> >  #undef __x86_get_cpu_features
> >>
> >> Why do we need both the conditional check and the function pointer hack?
> >
> > Because _dl_x86_init_cpu_features is called both indirectly and by IFUNC
> > reloc in dynamic executable, but it is only called by IFUNC reloc when
> > dlopen in static executable.
>
> I think we always need to call it eventually, as a dependency of filling
> in the cacheinfo data?

Yes.  _dl_x86_init_cpu_features is called twice by DL_PLATFORM_INIT
and IFUNC reloc in dynamic executable.  It is called once by IFUNC reloc
via dlopen in static executable.

> >> I expect that one of the function pointers can go, probably the one
> >> here.  The cache hierarchy data might be used by a string function that
> >> has not been selected by IFUNC.
> >>
> >
> > There are one IFUNC reloc in ld.so and the other in libc.so.  We need
> > both.
>
> libc.so should not need the relocation hack because we have
> __libc_early_init, which is also called after static dlopen and before
> constructors.
>

We want to call init_cacheinfo as early as possible.  __libc_early_init is
still too late.

-- 
H.J.