Re: [PATCH v3 3/5] gpio: gpiolib: Allow free() callback to be overridden

[Marc Zyngier]

On Thu, 12 May 2022 18:55:38 +0100,
"Lad, Prabhakar" <prabhakar.csengg@xxxxxxxxx> wrote:
> 
> Hi Marc,
> 
> On Thu, May 12, 2022 at 5:26 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >
> > On Thu, 12 May 2022 14:50:05 +0100,
> > "Lad, Prabhakar" <prabhakar.csengg@xxxxxxxxx> wrote:
> > >
> > > Hi Marc,
> > >
> > > On Thu, May 12, 2022 at 2:24 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, 12 May 2022 13:48:53 +0100,
> > > > "Lad, Prabhakar" <prabhakar.csengg@xxxxxxxxx> wrote:
> > > > >
> > > > > Hi Marc,
> > > > >
> > > > > Thank you for the review.
> > > > >
> > > > > On Thu, May 12, 2022 at 12:19 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Wed, 11 May 2022 19:32:08 +0100,
> > > > > > Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > Allow free() callback to be overridden from irq_domain_ops for
> > > > > > > hierarchical chips.
> > > > > > >
> > > > > > > This allows drivers to free any resources which are allocated during
> > > > > > > populate_parent_alloc_arg().
> > > > > >
> > > > > > Do you mean more than the fwspec? I don't see this being used.
> > > > > >
> > > > > The free callback is used in patch 5/5 where free is overridden by
> > > > > rzg2l_gpio_irq_domain_free. I just gave an example there as an
> > > > > populate_parent_alloc_arg()  In actual in the child_to_parent_hwirq
> > > > > callback I am using a bitmap [0] to get a free tint slot, this bitmap
> > > > > needs freeing up when the GPIO interrupt is released from the driver
> > > > > that as when overridden free callback frees the allocated tint slot so
> > > > > that its available for re-use.
> > > >
> > > > Right, so that's actually a different life-cycle, and the whole
> > > > populate_parent_alloc_arg() is a red herring. What you want is to free
> > > > resources that have been allocated via some other paths. It'd be good
> > > Is there any other path which I have missed where I can free up resources?
> >
> > No, that's the only one. It is just that usually, the alloc()
> > callback is where you are supposed to perform... allocations.
> >
> OK.
> 
> > It'd be good if you could move your allocation there, as I would
> > expect calls to child_to_parent_hwirq() to be idempotent.
> >
> For now I'll go with the current implementation, as currently a an
> array is maintained which is tied with the tint slot and child (which
> is obtained from child_to_parent_hwirq)
> 
> > >
> > > > if your commit message actually reflected this instead of using an
> > > > example that doesn't actually exist.
> > > >
> > > My bad, I will update the commit message.
> > >
> > > > >
> > > > > > There is also the question of why we need to have dynamic allocation
> > > > > > for the fwspec itself. Why isn't that a simple stack allocation in the
> > > > > > context of gpiochip_hierarchy_irq_domain_alloc()?
> > > > > >
> > > > > you mean gpio core itself should handle the fwspec
> > > > > allocation/freeing?
> > > >
> > > > Yes. The only reason we resort to dynamic allocation is because
> > > > ThunderX is using MSI-based GPIOs, and thus doesn't use a fwspec (no
> > > > firmware is involved here).
> > > >
> > > I see..
> > >
> > > > If we had a union of the two types, we could just have a stack
> > > > variable, and pass that along, completely sidestepping the whole
> > > > dynamic allocation/freeing business.
> > > >
> > > Right agreed.
> >
> > FWIW, I've just posted a PoC patch[1].
> >
> I guess I'll have to rebase my changes on top of it now ;)

Not yet. Let's see what people say about it.

	M.

-- 
Without deviation from the norm, progress is not possible.