Re: [Qemu-devel] [PATCH 4/4] iothread: push gcontext earlier in the thread_fn

[Marc-André Lureau]

Hi

On Mon, Mar 4, 2019 at 3:27 AM Peter Xu <address@hidden> wrote:
>
> On Fri, Mar 01, 2019 at 04:25:17PM +0000, Stefan Hajnoczi wrote:
> > On Thu, Feb 28, 2019 at 01:58:56PM +0800, Peter Xu wrote:
> > > On Wed, Feb 27, 2019 at 01:38:38PM +0000, Stefan Hajnoczi wrote:
> > > > On Fri, Feb 22, 2019 at 02:57:24PM +0800, Peter Xu wrote:
> > > > > On Fri, Feb 22, 2019 at 07:37:02AM +0100, Marc-André Lureau wrote:
> > > > > > Hi
> > > > > >
> > > > > > On Fri, Feb 22, 2019 at 4:14 AM Peter Xu <address@hidden> wrote:
> > > > > > >
> > > > > > > We were pushing the context until right before running the 
> > > > > > > gmainloop.
> > > > > > > Now since we have everything unconditionally, we can move this
> > > > > > > earlier.
> > > > > > >
> > > > > > > One benefit is that now it's done even before init_done_sem, so as
> > > > > > > long as the iothread user calls iothread_create() and completes, 
> > > > > > > we
> > > > > > > know that the thread stack is ready.
> > > > > > >
> > > > > >
> > > > > > This will change the default context in the iothread, for code 
> > > > > > running
> > > > > > there. This may not be a good idea. Until now, only sources 
> > > > > > dispatched
> > > > > > from iothread_get_g_main_context() would have default context
> > > > > > associated to it.
> > > > > >
> > > > > > I don't know if the current behaviour is intentional, but it has 
> > > > > > some
> > > > > > logic. With this change, you may create hidden races, by changing 
> > > > > > the
> > > > > > default context of sources to the iothread.
> > > > >
> > > > > Yes I agree that the behavior will be changed in this patch that even
> > > > > if the iothread user does not use the gcontext they'll also have the
> > > > > context set.  I would think it should be ok because IMHO events hooked
> > > > > onto the aio context should not depend on the gcontext, but indeed I'd
> > > > > like to get some confirmation from others, especially the block layer.
> > > >
> > > > I don't understand why Patch 4 is desirable.  The comment about
> > > > init_done_sem isn't clear to me but I also wondered the same thing as
> > > > Marc-André.
> > > >
> > > > Can you explain why we should apply this patch?
> > >
> > > Hi, Stefan,
> > >
> > > The patch 4 itself does not help much for current QEMU, but it should
> > > be required to replace the patch that Marc-Andre proposed below:
> > >
> > > https://lists.gnu.org/archive/html/qemu-devel/2019-02/msg05460.html [1]
> > >
> > > And IMHO patch 4 along with this whole series should be a cleaner
> > > approach comparing to the one proposed in [1].  Here if my
> > > understanding is correct the problem is that
> > > g_main_context_push_thread_default() is really designed to be called
> > > at the very beginning of a thread creation but not dynamically called
> > > during the execution of a thread (prove is that it even does not have
> > > any error to return when failed to acquire the context so the caller
> > > will never know if it failed! see [2] below), in that sense this patch
> > > 4 can be seen as a tiny cleanup too.
> > >
> > > g_main_context_push_thread_default (GMainContext *context)
> > > {
> > >   GQueue *stack;
> > >   gboolean acquired_context;
> > >
> > >   acquired_context = g_main_context_acquire (context);
> > >   g_return_if_fail (acquired_context);  <------------- [2]
> > >
> > >   ...
> > > }
> >
> > I see.  This explains why you want to call it early.  If you're worried
> > about that then there should also be a comment warning people that this
> > must happen first before anything implicitly uses the thread's
> > GMainContext.
>
> Sure, I can add a comment above g_main_context_push_thread_default()
> to emphasize why it's preferred at the entry.
>
> >
> > What about Marc-André's concern about the change in behavior?  Now this
> > thread is associated with the GMainContext that isn't processed at in
> > aio_poll().  Previously the default main context would be used.
>
> IIUC Paolo has answered this question (Message-ID:
> <address@hidden>) - if the block
> layer (or say, the explicit aio_poll in the iothread_run) does not use
> the GMainContext at all then it should affect nothing, and with that
> there should have no real functional change.
>

It's a bold claim though, iothread isn't used only by the block layer.
It's used also by colo in qemu tree, and I used to have a branch for
virgl rendering for example. There might be other experimental work or
usage I missed. Furthermore, even the block layer, and its various
dependencies, is hard to review thoroughly. But it looks like it
doesn't rely on glib main context, I agree. But I would get prepared
for some weird (difficult to debug) regressions eventually.