On 09/12/2020 14:53, Linus Walleij wrote:
On Wed, Dec 9, 2020 at 12:19 PM Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
On Wed, Dec 9, 2020 at 9:51 AM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
On Tue, Dec 8, 2020 at 3:07 PM Enrico Weigelt, metux IT consult <lkml@xxxxxxxxx> wrote:
What we need to understand is if your new usecase is an outlier
so it is simplest modeled by a "mock" irq_chip or we have to design
something new altogether like notifications on changes. I suspect
irq_chip would be best because all drivers using GPIOs for interrupts
are expecting interrupts, and it would be an enormous task to
change them all and really annoying to create a new mechanism
on the side.
I would expect the platform abstraction to actually be close enough
to a chained irqchip that it actually works: the notification should
come in via vring_interrupt(), which is a normal interrupt handler
that calls vq->vq.callback(), calling generic_handle_irq() (and
possibly chained_irq_enter()/chained_irq_exit() around it) like the
other gpio drivers do should just work here I think, and if it did
not, then I would expect this to be just a bug in the driver rather
than something missing in the gpio framework.
Performance/latency-wise that would also be strongly encouraged.
Tglx isn't super-happy about the chained interrupts at times, as they
can create really nasty bugs, but a pure IRQ in fastpath of some
kinde is preferable and intuitive either way.
In my opinion the problem here is that proposed patch somehow describes Front end, but
says nothing about Backend and overall design.
What is expected to be virtualized? whole GPIO chip? or set of GPIOs from different GPIO chips?
Most often nobody want to give Guest access to the whole GPIO chip, so, most probably, smth. similar to
GPIO Aggregator will be needed.
How is situation going to be resolved that GPIO framework and IRQ framework are independent, but intersect, and
GPIO controller drivers have no idea who and how requesting GPIO IRQ - threaded vs non-threaded vs oneshot.
So, some information need to be transferred to Back end - at minimum IRQ triggering type.
Overall, it might be better to start from pure gpio and leave GPIO IRQ aside.
--
Best regards,
grygorii
