Re: [PATCH] genirq/manage: Reduce priority of forced secondary IRQ handler

[Crystal Wood <crwood@redhat.com>]

On Fri, 2025-10-24 at 15:33 +0200, Sebastian Andrzej Siewior wrote:
> On 2025-10-03 13:25:53 [-0500], Crystal Wood wrote:
> > On Sun, 2025-09-21 at 15:12 +0200, Lukas Wunner wrote:
> > > On Sat, Sep 20, 2025 at 11:20:26PM +0200, Thomas Gleixner wrote:
> > > > I obviously understand that the proposed change squashs the whole class
> > > > of similar (not yet detected) issues, but that made me look at that
> > > > particular instance nevertheless.
> > > > 
> > > > All aer_irq() does is reading two PCI config words, writing one and then
> > > > sticking 64bytes into a KFIFO. All of that is hard interrupt safe. So
> > > > arguably this AER problem can be nicely solved by the below one-liner,
> > > > no?
> > > 
> > > The one-liner (which sets IRQF_NO_THREAD) was what Crystal originally
> > > proposed:
> > > 
> > > https://lore.kernel.org/r/20250902224441.368483-1-crwood@redhat.com/
> > 
> > So, is the plan to apply the original patch then?
> 
> Did we settle on something?
> I wasn't sure if you can mix IRQF_NO_THREAD with IRQF_ONESHOT for shared
> handlers. If that is a thing, we Crystal's original would do it.

Do you mean mixing IRQF_NO_THREAD on this irq (which should eliminate
the forced IRQF_ONESHOT) with another shared irq that still has
IRQF_ONESHOT?

I suspect it was a non-issue because of IRQCHIP_ONESHOT_SAFE disabling
the forced oneshot (the other irq was pciehp).  Given that these are
pcie-specific, do they ever get used without MSI (which sets
IRQCHIP_ONESHOT_SAFE)[1]?

The issue seems to be that the type of oneshot we want for forced
threading (unmask after the first user-supplied handler) is different
from what we want for explicit IRQF_ONESHOT (unmask after the last
user-supplied handler).  If we separated those, then the semantics
would better match non-RT, and we'd only need to care about mixing
when it comes to explicit IRQF_NOSHOT.

> Then there is the question if we want to go the "class" problem to
> ensure that one handler can preempt the other.  And maybe I should
> clean up few ones tglx pointed out that provide a primary handler for
> no reason…

Either way works for me, as long as we pick at least one :-)

-Crystal

[1] I realize that the answer to "has any hardware designer ever
done this weird and bad thing?" is usually yes. :-P