Re: [PATCH] irqchip/gic-v3: handle DOMAIN_BUS_ANY in gic_irq_domain_select

[Dmitry Baryshkov]

On Mon, 19 Feb 2024 at 18:37, Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On Mon, 19 Feb 2024 16:21:06 +0000,
> Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> wrote:
> >
> > On Mon, 19 Feb 2024 at 17:53, Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > >
> > > On Mon, 19 Feb 2024 14:47:37 +0000,
> > > Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> wrote:
> > > >
> > > > Before the commit de1ff306dcf4 ("genirq/irqdomain: Remove the param
> > > > count restriction from select()") the irq_find_matching_fwspec() was
> > > > handling the DOMAIN_BUS_ANY on its own. After this commit it is a job of
> > > > the select() callback. However the callback of GICv3 (even though it got
> > > > modified to handle zero param_count) wasn't prepared to return true for
> > > > DOMAIN_BUS_ANY bus_token.
> > > >
> > > > This breaks probing of any of the child IRQ domains, since
> > > > platform_irqchip_probe() uses irq_find_matching_host(par_np,
> > > > DOMAIN_BUS_ANY) to check for the presence of the parent IRQ domain.
> > > >
> > > > Fixes: 151378251004 ("irqchip/gic-v3: Make gic_irq_domain_select() robust for zero parameter count")
> > > > Fixes: de1ff306dcf4 ("genirq/irqdomain: Remove the param count restriction from select()")
> > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
> > > > ---
> > > >  drivers/irqchip/irq-gic-v3.c | 3 ++-
> > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c
> > > > index 6fb276504bcc..e9e9643c653f 100644
> > > > --- a/drivers/irqchip/irq-gic-v3.c
> > > > +++ b/drivers/irqchip/irq-gic-v3.c
> > > > @@ -1696,7 +1696,8 @@ static int gic_irq_domain_select(struct irq_domain *d,
> > > >
> > > >       /* Handle pure domain searches */
> > > >       if (!fwspec->param_count)
> > > > -             return d->bus_token == bus_token;
> > > > +             return d->bus_token == bus_token ||
> > > > +                     bus_token == DOMAIN_BUS_ANY;
> > > >
> > > >       /* If this is not DT, then we have a single domain */
> > > >       if (!is_of_node(fwspec->fwnode))
> > > >
> > >
> > > I really dislike the look of this. If that's the case, any irqchip
> > > that has a 'select' method (such as imx-intmux) should be similarly
> > > hacked. And at this point, this should be handled by the core code.
> > >
> > > Can you try this instead? I don't have any HW that relies on
> > > behaviour, but I'd expect this to work.
> > >
> > > Thanks,
> > >
> > >         M.
> > >
> > > diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
> > > index aeb41655d6de..3dd1c871e091 100644
> > > --- a/kernel/irq/irqdomain.c
> > > +++ b/kernel/irq/irqdomain.c
> > > @@ -449,7 +449,7 @@ struct irq_domain *irq_find_matching_fwspec(struct irq_fwspec *fwspec,
> > >          */
> > >         mutex_lock(&irq_domain_mutex);
> > >         list_for_each_entry(h, &irq_domain_list, link) {
> > > -               if (h->ops->select)
> > > +               if (h->ops->select && bus_token != DOMAIN_BUS_ANY)
> > >                         rc = h->ops->select(h, fwspec, bus_token);
> > >                 else if (h->ops->match)
> > >                         rc = h->ops->match(h, to_of_node(fwnode), bus_token);
> >
> > This works. But I wonder if the following change is even better. WDYT?
> >
> > diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
> > index aeb41655d6de..2f0d2700709e 100644
> > --- a/kernel/irq/irqdomain.c
> > +++ b/kernel/irq/irqdomain.c
> > @@ -449,14 +449,17 @@ struct irq_domain
> > *irq_find_matching_fwspec(struct irq_fwspec *fwspec,
> >          */
> >         mutex_lock(&irq_domain_mutex);
> >         list_for_each_entry(h, &irq_domain_list, link) {
> > -               if (h->ops->select)
> > +               if (fwnode != NULL &&
> > +                   h->fwnode == fwnode &&
> > +                   bus_token == DOMAIN_BUS_ANY)
> > +                       rc = true;
> > +               else if (h->ops->select)
> >                         rc = h->ops->select(h, fwspec, bus_token);
> >                 else if (h->ops->match)
> >                         rc = h->ops->match(h, to_of_node(fwnode), bus_token);
> >                 else
> >                         rc = ((fwnode != NULL) && (h->fwnode == fwnode) &&
> > -                             ((bus_token == DOMAIN_BUS_ANY) ||
> > -                              (h->bus_token == bus_token)));
> > +                             (h->bus_token == bus_token));
> >
> >                 if (rc) {
> >                         found = h;
> >
>
> Can't say I like it either. It duplicates the existing check without
> any obvious benefit. Honestly, this code is shit enough that we should
> try to make it simpler, not more complex...

Only the fwnode conditions are duplicated. And it makes sense: we
should check for the DOMAIN_BUS_ANY first, before going to select. I'm
not sure whether at some point we'd have to add (&& bus_token !=
DOMAIN_BUS_ANY) to the ops->match check.

>
> I'd rather we keep the impact as minimal as possible, and use the
> upcoming weeks to weed out the effects of these changes (there is
> another report of some Renesas machine falling over itself here[1]).
>
> Thanks,
>
>         M.
>
> [1] https://lore.kernel.org/all/170802702416.398.14922976721740218856.tip-bot2@tip-bot2
>
> --
> Without deviation from the norm, progress is not possible.



-- 
With best wishes
Dmitry