On Thu, Mar 31, 2022 at 05:44:55PM +0200, Arnd Bergmann wrote:
> On Thu, Mar 31, 2022 at 1:49 PM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > > +
> > > +static irqreturn_t a64fx_diag_handler(int irq, void *dev_id)
> > > +{
> > > + handle_sysrq('c');
> >
> >
> > Why is this calling this sysrq call? From an interrupt? Why?
> >
> > And you are hard-coding "c", are you sure?
>
> This is an actual sysrq driver in the traditional sense, where you can send
> a single interrupt to the machine from the outside over a side channel.
>
> I suggested sysrq instead of just panic() to make it a bit more flexible.
> Unfortunately there is no additional data, so it comes down to always
> sending the same character.
>
> It would be possible to make that character configurable with a module
> parameter or something like that, but I'm not sure that is an improvement.
> Maybe you have another idea for this.
Given the interrupt can be dismissed then offering non-fatal actions in
response the chassis command seems reasonable.
There is some prior art for this sort of feature. AFAICT SGI UV has a
similar mechanism that can send an NMI-with-no-side-channel to the
kernel. The corresponding driver offers a range of actions using a
module parameter:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/platform/uv/uv_nmi.c#n180
I don't think a hardcoded 'c' makes any sense. With a hardcoded argument
it is just obfuscation. However it is certainly seems attractive to be
able to reuse handle_sysrq() to provide a more powerful set of actions.
Daniel.
