[PATCH 08/14] tty/serial/kgdboc: Add and wire up clear_irqs callback

On Mon, Sep 10, 2012 at 08:19:18PM +0100, Alan Cox wrote:

On Mon, 10 Sep 2012 10:57:04 -0700
Anton Vorontsov [off-list ref] wrote:

quoted

On Mon, Sep 10, 2012 at 12:16:24PM +0100, Alan Cox wrote:

quoted

quoted

serial port, the CPU receives NMI exception, and we fall into KDB
shell.  So, it is our "debug console", and it is able to interrupt
(and thus debug) even IRQ handlers themselves.

You seem to have an assumption of single core here. What happens if
the NMI hits CPU #0 and the serial IRQ hits CPU #1 simultaneously ?

If you can't redirect all serial IRQs to NMI context, e.g. sometimes you get
NMIs, sometimes IRQs, then your NMI handling is not deterministic, and surely
this is not supported.

This seems like arch specific magic leaking into the tty code. Surely
this should be buried in the depths of the platform IRQ code ?

I'd not call this arch-specific. It's task-specific, and the task is quite
peculiar, true, but still nothing to do with arch.

If the arch is not able to turn serial IRQs into NMIs, then surely that arch is
not able to implement serial-triggered NMI/KDB-shell, so we don't support that
arch.

And we don't have any other way to accoplish the same: we need to clear the
serial interrupts, not some arch-specific interrupts. It is the serial device
that causes NMI, not some arch's magic. So we just provide the knob to manage
the device's functionality.

We do have a separate arch-specific NMI-masking knob that works at CPU level,
but if we mask NMI source at CPU level, we have to keep it always masked,
otherwise we'll keep reentering NMI. So, we really have to clear serial's
interrupt line.

That's why I came with clear_irqs callback, it has defined semantics for the
serial driver, and zero knowledge about any underlaying architecture or task.

Thanks,

Anton.