Re: [PATCH v2 1/4] genirq/irqdomain: Allow partial trimming of irq_data hierarchy

[Marc Zyngier]

On 2020-10-07 09:05, Marc Zyngier wrote:
> On 2020-10-06 21:39, Thomas Gleixner wrote:
> > On Tue, Oct 06 2020 at 11:11, Marc Zyngier wrote:
> > > It appears that some HW is ugly enough that not all the interrupts
> > > connected to a particular interrupt controller end up with the same
> > > hierarchy repth (some of them are terminated early). This leaves
> >
> >   depth?
> >
> > > the irqchip hacker with only two choices, both equally bad:
> > >
> > > - create discrete domain chains, one for each "hierarchy depth",
> > >   which is very hard to maintain
> > >
> > > - create fake hierarchy levels for the shallow paths, leading
> > >   to all kind of problems (what are the safe hwirq values for these
> > >   fake levels?)
> > >
> > > Instead, let's offer the possibility to cut short a single interrupt
> >
> > s/let's offer/implement/
>
> Thanks for that, I'll fix it locally.
>
> [...]
>
> > This is butt ugly, really. Especially the use case where the tegra PMC
> > domain removes itself from the hierarchy from .alloc()
>
> I don't disagree at all. It is both horrible and dangerous.
>
> My preference would have been to split the PMC domain into discrete
> domains, each one having having its own depth. But that's incredibly
> hard to express in DT, and would break the combination of old/new
> DT and kernel.
>
> > That said, I don't have a better idea either. Sigh...
>
> A (very minor) improvement would be to turn the trim call in the PMC 
> driver into
> a flag set in the first invalid irq_data structure, and let
> __irq_domain_alloc_irqs() do the dirty work.
>
> Still crap, but at least would prevent some form of abuse. Thoughts?

Actually, I wonder whether we can have a more general approach:

A partial hierarchy that doesn't have an irq_data->chip pointer 
populated
cannot be valid. So I wonder if the least ugly thing to do is to just 
drop
any messing about in the PMC driver, and instead to let 
__irq_domain_alloc_irqs()
do the culling, always, by looking for a NULL pointer in irq_data->chip.

Not any less ugly, but at least doesn't need any driver intervention.

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