Re: [PATCH 2/2] soundwire: bus: Allow SoundWire peripherals to register IRQ handlers

[Charles Keepax <ckeepax@opensource.cirrus.com>]

On Fri, Jan 20, 2023 at 10:20:50AM -0600, Pierre-Louis Bossart wrote:
> On 1/20/23 03:59, Charles Keepax wrote:
> > On Thu, Jan 19, 2023 at 11:12:04AM -0600, Pierre-Louis Bossart wrote:
> >> There should be an explanation and something checking that both are not
> >> used concurrently.
> > 
> > I will try to expand the explanation a litte, but I dont see any
> > reason to block calling both handlers, no ill effects would come
> > for a driver having both and it is useful if any soundwire
> > specific steps are needed that arn't on other control buses.
> 
> I think it's problematic if the peripheral tries to wake-up the manager
> from clock-stop with both an in-band wake (i.e. drive the data line
> high) and a separate GPIO-based interrupt. It's asking for trouble IMHO.
> We spent hours in the MIPI team to make sure there were no races between
> the manager-initiated restarts and peripheral-initiated restarts, adding
> a 3rd mechanism in the mix gives me a migraine already.

Apologies but I am struggling see why this has any bearing on
the case of a device that does both an in-band and out-of-band
wake. The code we are adding in this patch will only be called in the
in-band case. handle_nested_irq doesn't do any hardware magic or
schedule any threads, it just calls a function that was provided
when the client called request_threaded_irq. The only guarantee
of atomicity you have on the interrupt_callback is sdw_dev_lock
and that is being held across both calls after the patch.

Could you be a little more specific on what you mean by this
represents a 3rd mechanism, to me this isn't a new mechanism just
an extra callback? Say for example this patch added an
interrupt_callback_early to sdw_slave_ops that is called just
before interrupt_callback.

@@ -1681,6 +1681,9 @@ static int sdw_handle_slave_alerts(struct sdw_slave *slave)
                                struct device *dev = &slave->dev;
                                struct sdw_driver *drv = drv_to_sdw_driver(dev->driver);
 
+                         if (drv->ops && drv->ops->interrupt_callback_early)
+                                 drv->ops->interrupt_callback_early(slave);
+
                                if (drv->ops && drv->ops->interrupt_callback) {
                                        slave_intr.sdca_cascade = sdca_cascade;
                                        slave_intr.control_port = clear;

Would that similarly worry you? As in is it the client driver
writer dealing with 2 points of entry that worries you, or
something deeper relating to the IRQs?

Also if it helps I could go over in a little more detail how
the IRQs on our device works and why that means I would prefer
to have the option to use both. There are alternatives but they
arn't really as pretty.

Thanks,
Charles