[PATCH] libstdc++: Implementation of class strand

[Thomas Rodgers]

IIRC, the Asio implementation checks to see if the current thread already
holds a lock, and if so, adopts the current lock and enqueues any child
tasks that are created and processes them after the user-provided function
returns, to avoid deadlock in the case that Jonathan outlines here.

On Wed, Mar 22, 2023 at 7:32 AM Jonathan Wakely via Libstdc++ <
libstdc++@gcc.gnu.org> wrote:

> On Tue, 23 Mar 2021 at 16:39, Jonathan Wakely <jwakely@redhat.com> wrote:
>
> > On 12/03/21 13:21 +0100, Alessio G. B. via Libstdc++ wrote:
> > >I expanded the implementation of the class strand of the Networking
> > >TS. Essentially, I
> > >implemented a token system so each thread knows when it can execute;
> > >the system is organized
> > >with 2 integers moving as a clock.
> >
> > Thanks for this patch. I'm not sure when I'll have time to review it,
> > and it might not be in time for the upcoming GCC 11 release. But the
> > patch has been received and will get reviewed, thanks.
> >
>
> Well I didn't think it would take me two years, sorry about that :-(
>
> + template<typename _Func>
> + void
> + invoke(unsigned int token, _Func&& __f)
> + {
> +  std::unique_lock<std::mutex> __lock(_M_mutex);
> +
> +  _M_cv.wait(__lock,
> +     [token, next_token = _M_next_token]()
> +     { return token == next_token; });
> +
> +  try { decay_t<_Func>{std::forward<_Func>(__f)}(); }
> +  catch(...) { }
> +
> +  _M_next_token++;
> +
> +  __lock.unlock();
> +
> +  _M_cv.notify_all();
> + }
>
> It looks like this will run a user-provided function while holding the
> mutex lock. Won't that deadlock if the task added to the strand adds
> another task to the same strand? Is that forbidden by some requirement in
> the TS that I've forgotten?
>
>