Re: [Qemu-devel] [RFC] New Migration Protocol using Visitor Interface

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Oct 03, 2011 at 09:42:02AM -0500, Anthony Liguori wrote:
> On 10/03/2011 09:11 AM, Michael S. Tsirkin wrote:
> >On Mon, Oct 03, 2011 at 08:43:54AM -0500, Anthony Liguori wrote:
> >>>>visit_start_array(v, "entries", errp);
> >>>>for (int i = 0; i<   s->size; i++) {
> >>>>     visit_type_int(v, NULL,&s->entry[i], errp);
> >>>>}
> >>>>visit_end_array(v, errp);
> >>>
> >>>Sequences can encode structures not just arrays.
> >>>How would you encode this for example:
> >>>
> >>>SEQUENCE OF { VQN: INTEGER, SEQUENCE { OPTIONAL VECTOR: INTEGER}  }
> >>
> >>visit_start_array(v, "vqs", errp);
> >>for (i = 0; i<  s->n_vqs; i++) {
> >>     // Array elements never have a name, hence NULL name
> >>     visit_start_struct(v, "VirtQueue", NULL, errp);
> >>     visit_type_int(v,&s->vq[i].num, "vqn", errp);
> >>
> >>     // Given this sub-struct an arbitrary name.  It could also be anonymous.
> >>     visit_start_struct(v, "MsixInfo", "msix_info", errp);
> >>     if (s->vq[i].msix_enabled) {
> >>         visit_type_int(v,&s->vq[i].vector, "vector", errp);
> >
> >Why is this a pointer to vector, btw?
> 
> So you can write a single visit function that works for input or output.
> 
> Think of the simple case like:
> 
> void visit_simple_type(Visitor *v, SimpleType *t, const char *name, Error **errp)
> {
>    visit_start_struct(v, "SimpleType", name, errp);
>    visit_type_int(v, &t->a, "a", errp);
>    visit_type_int(v, &t->b, "b", errp);
>    visit_end_struct(v, errp);
> }

Okay, so this actually stores the pointer to the integer somewhere?
So what is the lifetime requirement for this memory?
For how long must it stay around?

> For complex types like Virtio, you need to do a bit more.  You
> wouldn't do a simple for () {} loop but instead use the Visitor list
> mechanism.  That would eliminate the need to have to marshal n_vqs.

Example?

> >
> >>     }
> >>     visit_end_struct(v, errp);
> >>
> >>     visit_end_struct(v, errp);
> >>}
> >>visit_end_array(v, errp);
> >>
> >>This would also generate JSON of:
> >>
> >>'vqs': [ { 'vqn': 2, 'msix_info': { 'vector': 3 } } ]
> >
> >How would optional fields be handled?
> 
> As far as the Visitor goes, if something is optional you just don't
> encode it. If you need to key off the presence of a field,
> presumably you could just check to see whether it succeeded or
> failed to visit that field.

It would typically depend on the value.


> I'm not 100% sure if you can do a
> single input/output visitor when you have optional fields.
> 
> My rough thinking is that each device would have a input/output
> visitor callback that took the same signature.  That gives the
> flexibility of having two separate interfaces but in the common
> case, you just pass the same function for both.
> 
> >Specifically
> >the case where first field in a sequence tells
> >you the meaning of the following ones?
> 
> Can you give me the example in ASN.1?
> 
> Regards,
> 
> Anthony Liguori

That would be a selection from CHOICE.
Note that CHOICE doesn't affect encoding on the wire:
BER just uses the underlying type.