Re: [Qemu-devel] Re: Extending qemu_irq for reset signals

[Blue Swirl]

On 8/15/07, Paul Brook <address@hidden> wrote:
> On Wednesday 15 August 2007, Blue Swirl wrote:
> > On 8/15/07, Paul Brook <address@hidden> wrote:
> > > On Wednesday 15 August 2007, Paul Brook wrote:
> > > > > Okay, more explaining. This is the case where I'd want to use the
> > > > > signal: DMA controller ("upstream") can reset the slave device (ESP
> > > > > or Lance). DMA controller is created first and I also want to
> > > > > allocate reset signals at that point. Later when ESP is created, it
> > > > > should be possible to put ESP reset function and opaque data to the
> > > > > signal given but this is not possible with current API. Currently the
> > > > > DMA data would be passed to qemu_allocate_irqs.
> > > >
> > > > Ah, I see. The problem here is that you've got a cyclic dependency. For
> > > > DMA operations the ESP is in charge, so it makes sense to create the
> > > > subservient DMA device first. For the reset signals the DMA controller
> > > > is in charge so ideally you create the ESP device first. Because the
> > > > DMA interface is most complicated, it's probably takes precedence.
> > > >
> > > > I think you need to modify or use sparc32_dma_set_reset_data to take a
> > > > qemu_irq rather than a callback and opaque argument. Alternatively you
> > > > can move things around a bit and have the sun4m code do something
> > > > similar. i.e. the ESP and lance devices return the reset lines, then
> > > > the sun4m code pokes into the DMA device state.
> > >
> > > Oh, or you can pass a pointer to a qemu_irq from the DMA to the ESP and
> > > have the ESP poke its reset object in there that way.
> >
> > That's what I had in mind. Should I just extend the API for example with
> > /* Change the callback function and/or data */
> > void qemu_irq_change(qemu_irq irq, qemu_irq_handler handler,
> >                      void *opaque);
>
> I'm not so keen on that. It breaks the nice property of having the IRQ lines
> owned by the receiver. I was thinking something like:
>
> struct DMAState {
>   qemu_irq device_reset;
> };
> qemu_irq *dma_init()
> {
>   struct DMAState * d = malloc();
>   return &d->device_reset;
> }
>
> void esp_init(qemu_irq *resetp)
> {
>   ESPState *d = malloc();
>   *reset = *qemu_irq_alloc(d, 1);
> }
>
> void sun4m_init()
> {
>   qemu_irq *p = dma_init();
>   esp_init(p);
> }

Yes, that would work. I wasn't too happy about the change function either.