Re: [PATCH] libstdc++/112351 - deal with __gthread_once failure during locale init

[Jonathan Wakely <jwakely@redhat.com>]

On Mon, 6 Nov 2023 at 12:52, Richard Biener <rguenther@suse.de> wrote:
>
> On Mon, 6 Nov 2023, Jonathan Wakely wrote:
>
> > On Mon, 6 Nov 2023 at 11:52, Richard Biener <rguenther@suse.de> wrote:
> > >
> > > The following makes the C++98 locale init path follow the way the
> > > C++11 performs initialization.  This way we deal with pthread_once
> > > failing, falling back to non-threadsafe initialization which, given we
> > > initialize from the library, should be serialized by the dynamic
> > > loader already.
> > >
> > > Bootstrapped and tested on x86_64-unknown-linux-gnu, OK for trunk?
> > > And GCC 13 branch?
> > >
> > > Thanks,
> > > Richard.
> > >
> > >         PR libstdc++/112351
> > > libstdc++-v3/
> > >         * src/c++98/locale.cc (locale::facet::_S_get_c_locale):
> > >         Always perform non-threadsafe init when threadsafe init
> > >         failed.
> > > ---
> > >  libstdc++-v3/src/c++98/locale.cc | 7 ++-----
> > >  1 file changed, 2 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/libstdc++-v3/src/c++98/locale.cc b/libstdc++-v3/src/c++98/locale.cc
> > > index d308140bab7..e9bec1db3b6 100644
> > > --- a/libstdc++-v3/src/c++98/locale.cc
> > > +++ b/libstdc++-v3/src/c++98/locale.cc
> > > @@ -216,12 +216,9 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> > >  #ifdef __GTHREADS
> > >      if (__gthread_active_p())
> > >        __gthread_once(&_S_once, _S_initialize_once);
> > > -    else
> > >  #endif
> > > -      {
> > > -       if (!_S_c_locale)
> > > -         _S_initialize_once();
> > > -      }
> > > +    if (__builtin_expect (!_S_c_locale, 0))
> > > +      _S_initialize_once();
> > >      return _S_c_locale;
> > >    }
> >
> >
> > I think this has a problem, which is handled correctly in
> > src/c++11/locale_init.cc by checking _S_classic inside the
> > _S_initialize_once function.
>
> We check _S_c_locale here (it's just a pointer) instead of in

It's a __locale_struct* for glibc and int* for BSD. Hopefully it's
always something pointer-ish so that works.

> _S_initialize_once (), so I think the code is equivalent, no?
>
> > If the first call to __gthread_once does fail then _S_once will not be
> > changed. We will fall through to calling _S_initialize_once directly
> > (which is not thread-safe) and set _S_c_locale.
> >
> > The next time we call _S_initialize, __gthread_once will try to run
> > again, and because _S_once was not changed, it might call
> > _S_initialize_once() again, which writes to _S_c_locale again
> > (possibly causing a data race).
>
> Ah, yeah, so in the C++11 path the check for !_S_classic in
> locale::_S_initialize is redundant.  But good spot.

Redundant, but it avoids a call to _S_initialize_once, so probably
worth it for performance.

>
> > I don't think the slightly different code in src/c++11/locale_init.cc
> > is different in order to handle __gthread_once failing, I think it's
> > different because the effects of locale::facet::_S_initialize_once()
> > and locale::_S_initialize_once() are different. One is safe to call
> > more than once, and the other isn't.
> >
> > I don't think we need to care about __gthread_once failing at all, do
> > we? There are no error conditions for pthread_once, it always returns
> > 0 (previous POSIX revisions said it could return EINVAL for an
> > uninitialized pthread_once_t but that can't happen here as it's
> > correctly initialized in src/c++11/locale.cc). Is the concern that it
> > can fail for non-posix thread models? (I didn't check if any of them
> > can actually fail)
>
> The concern is that there are actual products out that break with the
> new I/O initialization in libstdc++ for GCC13+ because they have bugs.
> It's easy enough to work around those by the proposed patch (plus
> correction for the above issue).  I suppose the comment in
> locale::_S_initialize_once holds as well for the C++98 path.
>
> The failure mode of the product is that it overrides pthread_once
> but does nothing (not even indicate failure) when its pthread_*
> override mechanism isn't initialized yet.  With libstdc++ from GCC13
> we now use pthread_once "too early" and fail to initialize the locale
> object.
>
> Adjusted patch below.
>
> OK after another round of testing?

OK, thanks.

>
> Thanks,
> Richard.
>
>
> From 4e3fa2f4426a5a10d189587b63e4d7298c347b01 Mon Sep 17 00:00:00 2001
> From: Richard Biener <rguenther@suse.de>
> Date: Mon, 6 Nov 2023 11:31:40 +0100
> Subject: [PATCH] libstdc++/112351 - deal with __gthread_once failure during
>  locale init
> To: gcc-patches@gcc.gnu.org
>
> The following makes the C++98 locale init path follow the way the
> C++11 performs initialization.  This way we deal with pthread_once
> failing, falling back to non-threadsafe initialization which, given we
> initialize from the library, should be serialized by the dynamic
> loader already.
>
>         PR libstdc++/112351
> libstdc++-v3/
>         * src/c++98/locale.cc (locale::facet::_S_initialize_once):
>         Check whether _S_c_locale is already initialized.
>         (locale::facet::_S_get_c_locale): Always perform non-threadsafe
>         init when threadsafe init failed.
> ---
>  libstdc++-v3/src/c++98/locale.cc | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/libstdc++-v3/src/c++98/locale.cc b/libstdc++-v3/src/c++98/locale.cc
> index d308140bab7..1ef0c394cd7 100644
> --- a/libstdc++-v3/src/c++98/locale.cc
> +++ b/libstdc++-v3/src/c++98/locale.cc
> @@ -206,6 +206,12 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>    void
>    locale::facet::_S_initialize_once()
>    {
> +    // Need to check this because we could get called once from
> +    // _S_get_c_locale() when the program is single-threaded, and then again
> +    // (via __gthread_once) when it's multi-threaded.
> +    if (_S_c_locale)
> +      return;
> +
>      // Initialize the underlying locale model.
>      _S_create_c_locale(_S_c_locale, _S_c_name);
>    }
> @@ -216,12 +222,9 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>  #ifdef __GTHREADS
>      if (__gthread_active_p())
>        __gthread_once(&_S_once, _S_initialize_once);
> -    else
>  #endif
> -      {
> -       if (!_S_c_locale)
> -         _S_initialize_once();
> -      }
> +    if (__builtin_expect (!_S_c_locale, 0))
> +      _S_initialize_once();
>      return _S_c_locale;
>    }
>
> --
> 2.35.3
>