Am 16.01.2020 um 14:51 hat Stefan Hajnoczi geschrieben:
> > +static void vu_set_unix_socket(Object *obj, const char *value,
> > +                                            Error **errp)
> > +{
> > +    VubDev *vus = VHOST_USER_SERVER(obj);;
> > +
> > +    if (vus->unix_socket) {
> > +        error_setg(errp, "unix_socket property already set");
> > +        return;
> > +    }
> > +
> > +    vus->unix_socket = g_strdup(value);
> > +    vhost_user_server_start(vus, value, vus->name,
> > +                            vus->writable, errp);
> 
> Property setters should only perform input validation and store the
> data.  Actions like creating network connections, opening files, etc
> should happen later in a UserCreatableClass->complete() callback.
> 
> This is necessary because vus->writable is also a property and may be
> set after unix_socket.  The ->complete() callback is called after all
> setters so it can access the final values of all properties.
> 
> See iothread_class_init() and iothread_complete() for an example.

Ah, right, this is the correct way. I forgot about the existence of
.complete(), so please ignore what I wrote.

> > diff --git a/vl.c b/vl.c
> > index 86474a55c9..72ac506342 100644
> > --- a/vl.c
> > +++ b/vl.c
> > @@ -2553,6 +2553,10 @@ static bool object_create_initial(const char *type, QemuOpts *opts)
> >      }
> >  #endif
> > 
> > +    /* Reason: vhost-user-server property "name" */
> > +    if (g_str_equal(type, "vhost-user-server")) {
> > +        return false;
> > +    }
> 
> I don't understand why the "name" property introduces a creation order
> dependency.  It's just a string and has no dependency on other
> command-line objects.  Can you explain why this change is necessary?

I was confused at first, too, but it's just a naming problem: "name"
is what points to the block device to be exported. It should really be
"node-name".

> > +struct VuClient {
> > +    VuDev parent;
> > +    int refcount;
> > +    VubDev *blk;
> > +    QIOChannelSocket *sioc; /* The underlying data channel */
> > +    QIOChannel *ioc; /* The current I/O channel */
> > +    QTAILQ_ENTRY(VuClient) next;
> > +    bool closed;
> > +};
> > +VubDev *vub_dev_find(const char *name);
> 
> All -object instances already have an id=ID property.  There is no need
> to declare a separate "name" property.  Please look at iothread_by_id()
> and iothread_get_id() for examples.

It's questionable to me if this function is needed at all. It is only
used in vhost_user_server_start() to make sure that you don't export a
node twice - but what's even the problem with exporting it twice?

Kevin