Re: [PATCH 20/35] irqchip/gic-v4.1: Allow direct invalidation of VLPIs

[Marc Zyngier]

On 2019-09-28 03:02, Zenghui Yu wrote:
> On 2019/9/24 2:25, Marc Zyngier wrote:
> > Just like for INVALL, GICv4.1 has grown a VPE-aware INVLPI register.
> > Let's plumb it in and make use of the DirectLPI code in that case.
> > Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx>
> > ---
> >   drivers/irqchip/irq-gic-v3-its.c   | 19 +++++++++++++++++--
> >   include/linux/irqchip/arm-gic-v3.h |  1 +
> >   2 files changed, 18 insertions(+), 2 deletions(-)
> > diff --git a/drivers/irqchip/irq-gic-v3-its.c 
> > b/drivers/irqchip/irq-gic-v3-its.c
> > index b791c9beddf2..34595a7fcccb 100644
> > --- a/drivers/irqchip/irq-gic-v3-its.c
> > +++ b/drivers/irqchip/irq-gic-v3-its.c
> > @@ -1200,13 +1200,27 @@ static void wait_for_syncr(void __iomem 
> > *rdbase)
> >   static void direct_lpi_inv(struct irq_data *d)
> >   {
> > +	struct its_vlpi_map *map = get_vlpi_map(d);
> >   	struct its_collection *col;
> >   	void __iomem *rdbase;
> > +	u64 val;
> > +
> > +	if (map) {
> > +		struct its_device *its_dev = irq_data_get_irq_chip_data(d);
> > +
> > +		WARN_ON(!is_v4_1(its_dev->its));
> > +
> > +		val  = GICR_INVLPIR_V;
> > +		val |= FIELD_PREP(GICR_INVLPIR_VPEID, map->vpe->vpe_id);
> > +		val |= FIELD_PREP(GICR_INVLPIR_INTID, map->vintid);
> > +	} else {
> > +		val = d->hwirq;
> > +	}
> >
> >   	/* Target the redistributor this LPI is currently routed to */
> >   	col = irq_to_col(d);
>
> I think irq_to_col() may not work when GICv4.1 VLPIs are involved in.
>
> irq_to_col() gives us col_map[event] as the target redistributor,
> but the correct one for VLPIs should be 
> vlpi_maps[event]->vpe->col_idx.
> These two are not always pointing to the same physical RD.
> For example, if guest issues a MOVI against a VLPI, we will update 
> the
> corresponding vlpi_map->vpe and issue a VMOVI on ITS... but leave the
> col_map[event] unchanged.
>
> col_map[event] usually describes the physical LPI's CPU affinity, but
> when this physical LPI serves as something which the VLPI is backed 
> by,
> we take really little care of it.  Did I miss something here?

You didn't miss anything, and this is indeed another pretty bad bug.
The collection mapping is completely unused when the LPI becomes a
VLPI, and it is only the vpe->col_id that matters (which gets updated
on VMOVP).

This shows that irq_to_col() is the wrong abstraction, and what we're
interested is something that is more like 'irq to cpuid', allowing us
to directly point to the right distributor.

Please see the patch I just pushed[1], which does that.

Thanks,

        M.

[1] 
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/commit/?h=irq/gic-v4.1-devel&id=aff363113eb26b6840136b69c2c7db2ea691db20
--
Jazz is not dead. It just smells funny...