[PATCH] irqchip/gic-v3-its: Flush GICR caching for a cross node collection move of an irq

[marc.zyngier@arm.com (Marc Zyngier)]

On 20/12/17 09:34, Ganapatrao Kulkarni wrote:
> Hi Marc,
> 
> On Wed, Dec 20, 2017 at 2:56 PM, Marc Zyngier <marc.zyngier@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 (no longer valid) mapping of the interrupt.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.kulkarni@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 might still have cached data for a previous (no longer valid) mapping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."

Doing some documentation archaeology, I found that this wording has been
dropped from the engineering specification in August 2014, and was never
included in the architecture specification. I suggest you start using a
slightly more up-to-date set of documentation...

Now, back to your point: what it says in the bit of (confidential)
document that you quoted is that the *HW* must perform the invalidation
(that's what the words "implementations" and "ITS" refer to), not some
random bits of SW.

If you know of an implementation that suffers from this, please resend a
patch that handles this as a quirk, with a proper erratum number.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...