Hi Marc,
On Wed, Dec 20, 2017 at 2:56 PM, Marc Zyngier <marc.zyng...@arm.com> wrote:
> On 20/12/17 09:15, Ganapatrao Kulkarni wrote:
>> When an interrupt is moved, it is possible that an implementation that
>> supports caching might still have cached data for a previous
>> (no longer valid) mapping of the interrupt. In particular, in a distributed
>> GIC implementation like multi-socket SoC platfroms. Hence it is necessary
>> to flush cached entries after cross node collection migration.
>>
>> Signed-off-by: Ganapatrao Kulkarni <ganapatrao.kulka...@cavium.com>
>> ---
>> drivers/irqchip/irq-gic-v3-its.c | 6 ++++++
>> 1 file changed, 6 insertions(+)
>>
>> diff --git a/drivers/irqchip/irq-gic-v3-its.c
>> b/drivers/irqchip/irq-gic-v3-its.c
>> index 4039e64..ea849a1 100644
>> --- a/drivers/irqchip/irq-gic-v3-its.c
>> +++ b/drivers/irqchip/irq-gic-v3-its.c
>> @@ -1119,6 +1119,12 @@ static int its_set_affinity(struct irq_data *d, const
>> struct cpumask *mask_val,
>> if (cpu != its_dev->event_map.col_map[id]) {
>> target_col = &its_dev->its->collections[cpu];
>> its_send_movi(its_dev, target_col, id);
>> + /* Issue INV for cross node collection move on
>> + * multi socket systems.
>> + */
>> + if (cpu_to_node(cpu) !=
>> + cpu_to_node(its_dev->event_map.col_map[id]))
>> + its_send_inv(its_dev, id);
>> its_dev->event_map.col_map[id] = cpu;
>> irq_data_update_effective_affinity(d, cpumask_of(cpu));
>> }
>>
>
> The MOVI command doesn't have any such requirement (it only mandates
> synchronization), and doesn't say anything about distributed vs monolithic.
GIC-v3 spec do mention to issue ITS INV command or a write to GICR_INVLPIR.
pasting below snippet of MOVI command description.
"When an interrupt is moved to a collection, it is possible that an
implementation that supports speculative caching
might still have cached data for a previous (no longer valid) mapping
of the interrupt. Hence, implementations
must take care to invalidate any data associated with an interrupt
when it is moved. In particular, in a distributed
implementation, the ITS must write to the appropriate GICR_* register
to perform the invalidation in the redistributor."
>
> What am I missing?
>
> M.
> --
> Jazz is not dead. It just smells funny...
thanks
Ganapat
