Re: [Qemu-devel] [PATCH v2 22/23] vhost-user: wait until backend init is completed

[Yuanhan Liu]

On Thu, Jun 30, 2016 at 05:22:43AM -0400, Marc-André Lureau wrote:
> Hi
> 
> ----- Original Message -----
> > On Fri, Jun 24, 2016 at 03:51:09PM +0200, address@hidden wrote:
> > > From: Marc-André Lureau <address@hidden>
> > > 
> > > The chardev waits for an initial connection before starting qemu,
> > > vhost-user wants the backend negotiation to be completed. vhost-user is
> > > started in the net_vhost_user_event callback, which is synchronously
> > > called after the socket is connected.
> > > 
> > > Signed-off-by: Marc-André Lureau <address@hidden>
> > > ---
> > >  net/vhost-user.c | 12 +++++++++++-
> > >  1 file changed, 11 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/net/vhost-user.c b/net/vhost-user.c
> > > index 95ed2d2..4badd9e 100644
> > > --- a/net/vhost-user.c
> > > +++ b/net/vhost-user.c
> > > @@ -24,6 +24,7 @@ typedef struct VhostUserState {
> > >      VHostNetState *vhost_net;
> > >      int watch;
> > >      uint64_t acked_features;
> > > +    bool started;
> > >  } VhostUserState;
> > >  
> > >  typedef struct VhostUserChardevProps {
> > > @@ -211,6 +212,7 @@ static void net_vhost_user_event(void *opaque, int
> > > event)
> > >              return;
> > >          }
> > >          qmp_set_link(name, true, &err);
> > > +        s->started = true;
> > >          break;
> > >      case CHR_EVENT_CLOSED:
> > >          qmp_set_link(name, false, &err);
> > > @@ -248,7 +250,15 @@ static int net_vhost_user_init(NetClientState *peer,
> > > const char *device,
> > >          s->chr = chr;
> > >      }
> > >  
> > > -    qemu_chr_add_handlers(chr, NULL, NULL, net_vhost_user_event,
> > > nc[0].name);
> > > +    do {
> > > +        Error *err = NULL;
> > > +        if (qemu_chr_wait_connected(chr, &err) < 0) {
> > > +            error_report_err(err);
> > > +            return -1;
> > > +        }
> > > +        qemu_chr_add_handlers(chr, NULL, NULL,
> > > +                              net_vhost_user_event, nc[0].name);
> > > +    } while (!s->started);
> > 
> > 
> > I haven't looked at your patchset carefully yet, but I did a quick test,
> > and showed that above code piece just breaks vhost-user: it's a dead loop
> > that vhost-user net will be initiated again and again.
> 
> Interesting, thanks a lot for testing!
> 
> The vhost-user-test works just fine, as well as vhost-user-bridge.
> 
> > The dead loop is due to, we check "s->started" corresponding to last nc,
> > while we set "s->started" corresponding to the first nc.
> 
> I see, that makes sense with multiqueue, I'll update the code. It would be 
> really nice to have a basic multiqueue test in vhost-user-test.

My bad. It was actually my task to add the multiqueue test: it was
postponed, or even forgotten :(. If you don't mind, feel free to
add it. If you have not time for that, I will spend some time to
fix what I left.

        --yliu