Re: [Qemu-devel] [RFC v5 21/26] qmp: isolate responses into io thread

[Peter Xu <peterx@redhat.com>]

On Mon, Dec 18, 2017 at 01:52:17PM +0800, Peter Xu wrote:
> On Thu, Dec 14, 2017 at 01:43:59PM +0000, Stefan Hajnoczi wrote:
> > On Tue, Dec 05, 2017 at 01:51:55PM +0800, Peter Xu wrote:
> > > @@ -4429,6 +4515,13 @@ void monitor_cleanup(void)
> > >       */
> > >      iothread_stop(mon_global.mon_iothread);
> > >  
> > > +    /*
> > > +     * After we have IOThread to send responses, it's possible that
> > > +     * when we stop the IOThread there are still replies queued in the
> > > +     * responder queue.  Flush all of them.
> > > +     */
> > > +    monitor_qmp_bh_responder(NULL);
> > 
> > This doesn't work because monitor_qmp_bh_responder() does not guarantee
> > that the full response has been written when it returns.
> > 
> > When qemu_chr_fe_write() returns EAGAIN then qemu_chr_fe_add_watch() is
> > used to register an event loop callback when the chardev becomes
> > writable again.  But you stopped the event loop using iothread_stop() so
> > we will never complete the write.
> 
> Good catch...
> 
> Actually I just noticed that for char frontend I missed a place to use
> the chardev context for polling.  So before the flushing I possibly
> need this:
> 
> diff --git a/chardev/char-fe.c b/chardev/char-fe.c                             
> index ee6d596100..462c529f19 100644    
> --- a/chardev/char-fe.c                
> +++ b/chardev/char-fe.c                
> @@ -356,7 +356,7 @@ guint qemu_chr_fe_add_watch(CharBackend *be, GIOCondition cond,                                                                           
>      }                                 
>                                        
>      g_source_set_callback(src, (GSourceFunc)func, user_data, NULL);           
> -    tag = g_source_attach(src, NULL); 
> +    tag = g_source_attach(src, be->chr->gcontext);                            
>      g_source_unref(src);              
>                                        
>      return tag;                       
> 
> Otherwise it'll be still be run in main thread always.
> 
> (I guess I haven't yet encountered an EAGAIN for it so far)
> 
> > 
> > I suggest draining the monitor while the IOThread is still running
> > (that way the AioContext and GMainContext are still operational).  You
> > can:
> > 1. Suspend the monitor so new commands will not be read.
> > 2. Wait until all responses and outbuf are empty.
> > 
> > Another option is moving the chardev back to the main loop but I'm not
> > sure if the chardev subsystem supports that.
> 
> Your suggestion is good to me.  I'll do that in IOThread before it
> stops.  Thanks!

Hmm, after a second thought, I think maybe it would be nicer to just
call monitor_flush() after the monitor iothread is stopped.

Firstly, that is perfectly legal - since AFAICT commit 6cff3e8594
("monitor: protect outbuf and mux_out with mutex", 2014-06-23) is
tailored for doing that, though that was for block iothread, not
monitor.

Then, it's still possible that the write buffer is full and we got
EAGAIN during flushing.  Then we'll plug another task into the monitor
gcontext which will never run since the iothread is stopped already.
IMHO that's totally fine - I think it's already our best effort to
flush the out buffer once at this point.  If the buffer is still full
after one attemp of monitor_flush(), we should just give up the 2nd
flushing instead of waiting here.  After all, we are quitting QEMU.

As a summary, my planned change of this issue would be quite
straightforward: call monitor_flush() when looping over monitors,
before destroying them.  And when destroying, if there is dangling
Monitor.out_watch, detach them.  Hope that works.  Thanks,

-- 
Peter Xu