[systemd-devel] EXT: sdbus_event loop state mark as volatile?

[Lennart Poettering]

On Fr, 23.04.21 11:24, Stephen Hemminger (stephen at networkplumber.org) wrote:

> On Fri, 6 Sep 2019 16:04:33 +0100
> Simon McVittie <smcv at collabora.com> wrote:
>
> > On Fri, 06 Sep 2019 at 06:57:22 +0000, Ray, Ian (GE Healthcare) wrote:
> > > If thread-safety is a design goal (and I don’t believe that it is [1])
> > > then atomic or thread-safe primitives should be used.
> > >
> > > [1] https://lists.freedesktop.org/archives/systemd-devel/2017-March/038519.html
> >
> > [1] is about sd-bus, not sd-event, and doesn't say anything about whether
> > sd-event is designed to be thread-safe or not.
> >
> > However, I think you're correct to say that struct sd_event is also only
> > designed to be used from the single thread that "owns" it.
> >
> > If you need a thread-safe event loop, then you need something like
> > GLib's GMainContext, with mutexes to protect its data structures against
> > concurrent access, and a well-defined mechanism for one main-context to
> > "post" events to other main-contexts (which might be running concurrently
> > in a different thread). Many other event loops are available; GMainContext
> > happens to be the one I'm most familiar with, and I know that it is
> > designed to be thread-safe.
> >
> > The price that things like GMainContext pay for being thread-safe is
> > that they are more complex and less efficient than sd-event: in general,
> > all operations on a thread-aware event loop have to pay the complexity
> > and performance cost of being thread-aware, even if the current program
> > only has one thread.
> >
> >     smcv
>
> Excuse me for reviving an old thread. But I see similar problem today
> (especially on Arm). The sd-event model uses signals so it is inherently
> subject to thread issues.

Hmm? sd-event uses signals only via signalfd(), so that it can
dispatch them as part of regular event loop handling. It doesn't
install a signal handler or anything like that.

> It looks like a stronger memory model is needed here (not volatile).
> Other projects use __atomic builtins for this.

All of sd-event's data structures should be accessed from a single
thread only, in a single non-signal execution context.

Lennart

--
Lennart Poettering, Berlin