On Tue, Sep 04, 2018 at 04:30:12PM +0200, Andrew Jones wrote:
> On Tue, Sep 04, 2018 at 03:29:26PM +0200, Christoffer Dall wrote:
> > On Tue, Sep 04, 2018 at 10:13:36AM +0200, Andrew Jones wrote:
> > > We need to support at least two redistributor regions in order to
> > > support more than 123 vcpus (we select 8 because that should be
> > > plenty). Also bump NR_CPUS to 512, since that's what KVM currently
> > > supports.
> > >
> > > Signed-off-by: Andrew Jones <drjones@xxxxxxxxxx>
> > > ---
> > > v2:
> > > - Peter asked if it would be neater to be more general. At first
> > > I interpreted that as better, i.e. more general, but immediately
> > > after starting the approach I saw how much _neater_ the code is.
> > > So this version is indeed neater code and better (more general).
> > > Thanks Peter!
> > >
> > >
> > > lib/arm/asm/gic-v3.h | 3 +++
> > > lib/arm/asm/setup.h | 2 +-
> > > lib/arm/gic-v3.c | 23 +++++++++++++----------
> > > lib/arm/gic.c | 14 +++++++++-----
> > > 4 files changed, 26 insertions(+), 16 deletions(-)
> > >
> > > diff --git a/lib/arm/asm/gic-v3.h b/lib/arm/asm/gic-v3.h
> > > index 1dceb9541f62..347be2f9da17 100644
> > > --- a/lib/arm/asm/gic-v3.h
> > > +++ b/lib/arm/asm/gic-v3.h
> > > @@ -49,8 +49,11 @@
> > > #include <asm/smp.h>
> > > #include <asm/io.h>
> > >
> > > +#define GICV3_NR_REDISTS 8
> > > +
> > > struct gicv3_data {
> > > void *dist_base;
> > > + void *redist_bases[GICV3_NR_REDISTS];
> > > void *redist_base[NR_CPUS];
> > > unsigned int irq_nr;
> > > };
> > > diff --git a/lib/arm/asm/setup.h b/lib/arm/asm/setup.h
> > > index b57ea13b9dd2..3215814603e4 100644
> > > --- a/lib/arm/asm/setup.h
> > > +++ b/lib/arm/asm/setup.h
> > > @@ -9,7 +9,7 @@
> > > #include <asm/page.h>
> > > #include <asm/pgtable-hwdef.h>
> > >
> > > -#define NR_CPUS 255
> > > +#define NR_CPUS 512
> > > extern u64 cpus[NR_CPUS]; /* per-cpu IDs (MPIDRs) */
> > > extern int nr_cpus;
> > >
> > > diff --git a/lib/arm/gic-v3.c b/lib/arm/gic-v3.c
> > > index 9b3fa5730478..7f41e0a1da01 100644
> > > --- a/lib/arm/gic-v3.c
> > > +++ b/lib/arm/gic-v3.c
> > > @@ -9,17 +9,20 @@
> > > void gicv3_set_redist_base(size_t stride)
> > > {
> > > u32 aff = mpidr_compress(get_mpidr());
> > > - void *ptr = gicv3_data.redist_base[0];
> > > u64 typer;
> > > -
> > > - do {
> > > - typer = gicv3_read_typer(ptr + GICR_TYPER);
> > > - if ((typer >> 32) == aff) {
> > > - gicv3_redist_base() = ptr;
> > > - return;
> > > - }
> > > - ptr += stride; /* skip RD_base, SGI_base, etc. */
> > > - } while (!(typer & GICR_TYPER_LAST));
> > > + int i = 0;
> > > +
> > > + while (gicv3_data.redist_bases[i]) {
> >
> > I don't get this. We never change I?
> >
> > (conincidentally, my test of gicv3-ipi never completes, I wonder if
> > these concepts are related...)
>
> Darn it. I dropped the '++i' I had in v1. My regression testing naturally
> only needs i=0 (I don't have anything with > 123 cpus to test on), so the
> problem didn't show up for me.
>
> Maybe you can add a ++i in and see if it works for you before I send v3?
> If it's still busted I'll have to stop shotgunning code, and maybe even
> find a way to test it myself.
>
Indeed, works perfectly. With an added i++:
Tested-by: Christoffer Dall <christoffer.dall@xxxxxxx>
Thanks,
Christoffer
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
