On Fri, Dec 09, 2016 at 04:08:00PM +0000, Andre Przywara wrote:
> On 08/12/16 17:50, Andrew Jones wrote:
> > +u32 gicv3_iar_irqnr(u32 iar)
> > +{
> > + return iar;
>
> I am probably a bit paranoid here, but the spec says that the interrupt
> ID is in bits[23:0] only (at most).
Indeed, I'll add '& ((1 << 24) - 1' here.
>
> > +}
> > +
> > +void gicv3_ipi_send_mask(int irq, const cpumask_t *dest)
> > +{
> > + u16 tlist;
> > + int cpu;
> > +
> > + assert(irq < 16);
> > +
> > + /*
> > + * For each cpu in the mask collect its peers, which are also in
> > + * the mask, in order to form target lists.
> > + */
> > + for_each_cpu(cpu, dest) {
> > + u64 mpidr = cpus[cpu], sgi1r;
> > + u64 cluster_id;
> > +
> > + /*
> > + * GICv3 can send IPIs to up 16 peer cpus with a single
> > + * write to ICC_SGI1R_EL1 (using the target list). Peers
> > + * are cpus that have nearly identical MPIDRs, the only
> > + * difference being Aff0. The matching upper affinity
> > + * levels form the cluster ID.
> > + */
> > + cluster_id = mpidr & ~0xffUL;
> > + tlist = 0;
> > +
> > + /*
> > + * Sort of open code for_each_cpu in order to have a
> > + * nested for_each_cpu loop.
> > + */
> > + while (cpu < nr_cpus) {
> > + if ((mpidr & 0xff) >= 16) {
> > + printf("cpu%d MPIDR:aff0 is %d (>= 16)!\n",
> > + cpu, (int)(mpidr & 0xff));
> > + break;
> > + }
> > +
> > + tlist |= 1 << (mpidr & 0xf);
> > +
> > + cpu = cpumask_next(cpu, dest);
> > + if (cpu >= nr_cpus)
> > + break;
> > +
> > + mpidr = cpus[cpu];
> > +
> > + if (cluster_id != (mpidr & ~0xffUL)) {
> > + /*
> > + * The next cpu isn't in our cluster. Roll
> > + * back the cpu index allowing the outer
> > + * for_each_cpu to find it again with
> > + * cpumask_next
> > + */
> > + --cpu;
> > + break;
> > + }
> > + }
> > +
> > + /* Send the IPIs for the target list of this cluster */
> > + sgi1r = (MPIDR_TO_SGI_AFFINITY(cluster_id, 3) |
> > + MPIDR_TO_SGI_AFFINITY(cluster_id, 2) |
> > + irq << 24 |
> > + MPIDR_TO_SGI_AFFINITY(cluster_id, 1) |
> > + tlist);
> > +
> > + gicv3_write_sgi1r(sgi1r);
> > + }
> > +
> > + /* Force the above writes to ICC_SGI1R_EL1 to be executed */
> > + isb();
> > +}
>
> Wow, this is really heavy stuff, especially for a Friday afternoon ;-)
> But I convinced myself that it's correct. The only issue is that it's
> sub-optimal if the MPIDRs of the VCPUs are not in order, say: 0x000,
> 0x100, 0x001.
> In this case we do three register writes instead of the minimal two.
> But it's still correct, so it's actually a minor nit just to prove that
> I checked the algorithm ;-)
>
> So apart from the minor comment above:
>
> Reviewed-by: Andre Przywara <andre.przywara@xxxxxxx>
Thanks!
drew
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
