[PATCH 3/7] firmware: Add RNMI/RNME handler infrastructure

[Anup Patel]

On Fri, Apr 17, 2026 at 1:56 AM Evgeny Voevodin
<evvoevod at tenstorrent.com> wrote:
>
> On Fri, Apr 10, 2026 at 1:09 AM Anup Patel <anup at brainfault.org> wrote:
> > > > Drop these platform callbacks from this patch. We will be
> > > > extending sbi_irqchip to support NMIs. For NME, we will
> > > > can simply crash.
> > >
> > > Agreed. I'll remove the platform callbacks. For RNMI, I'll
> > > route through sbi_irqchip once that support is in place. For
> > > RNME, I'll simplify to just print diagnostics and hang.
>
> Hi Anup,
>
> I've been rebasing and reworking the Smrnmi series based on our
> discussion. I can remove the platform runtime callbacks
> (ops->rnmi_handler / ops->rnme_handler) as agreed, and I can still
> have the initial Smrnmi infrastructure in place: platforms install
> trap handlers for RNMI/RNME vectors, detect the extension from
> device tree, and initialize MNSCRATCH.
>
> However, without irqchip NMI support there's no way for platforms
> to actually handle NMIs once they fire. The default RNMI handler
> would just log and hang. This means NMIs would need to stay
> disabled (NMIE=0) until that support lands, even on platforms that
> are otherwise ready for Smrnmi.
>
> I couldn't find ongoing work on irqchip NMI support, so I wanted
> to check how you'd like to proceed:
>
> 1. Ship the Smrnmi infrastructure with NMIs effectively disabled
>    (log and hang). Platforms carry local patches for NMI handling
>    until irqchip NMI support lands.
>
> 2. Keep ops->rnmi_handler as a temporary bridge, marked for
>    migration to irqchip.
>
> 3. Include a minimal irqchip NMI extension in this series: a
>    standalone registration API (sbi_irqchip_register_nmi_handler /
>    sbi_irqchip_process_nmi) with a simple static callback -- not
>    tied to a specific irqchip device since NMI is a system-level
>    event. Platforms would register their NMI handler during
>    irqchip_init, and sbi_trap_rnmi_handler would dispatch
>    through it.
>

I am leaning towards option #2 and as a separate series
the sbi_irqchip can be extended to support NMIs.

Regards,
Anup