On Sun, Feb 10, 2019 at 07:49:12PM +0100, Thomas Gleixner wrote:
> On Fri, 25 Jan 2019, Ming Lei wrote:
> > +static int nvme_setup_affinity(const struct irq_affinity *affd,
> > + struct irq_affinity_desc *masks,
> > + unsigned int nmasks)
> > +{
> > + struct nvme_dev *dev = affd->priv;
> > + int affvecs = nmasks - affd->pre_vectors - affd->post_vectors;
> > + int curvec, usedvecs;
> > + int i;
> > +
> > + nvme_calc_io_queues(dev, nmasks);
>
> So this is the only NVME specific information. Everything else can be done
> in generic code. So what you really want is:
>
> struct affd {
> ...
> + calc_sets(struct affd *, unsigned int nvecs);
> ...
> }
>
> And sets want to be actually inside of the affinity descriptor structure:
>
> unsigned int num_sets;
> unsigned int set_vectors[MAX_SETS];
>
> We surely can define a sensible maximum of sets for now. If that ever turns
> out to be insufficient, then struct affd might become to large for the
> stack, but for now, using e.g. 8, there is no need to do so.
>
> So then the logic in the generic code becomes exactly the same as what you
> added to nvme_setup_affinity():
>
> if (affd->calc_sets) {
> affd->calc_sets(affd, nvecs);
> } else if (!affd->num_sets) {
> affd->num_sets = 1;
> affd->set_vectors[0] = affvecs;
> }
>
> for (i = 0; i < affd->num_sets; i++) {
> ....
> }
>
> See?
OK, will do this way in V2, then we can avoid drivers to abuse
the callback.
Thanks,
Ming
