> Jean, Wolfram, Benjamin, or someone else, could you please check Thomas'
> questions above and let us know what you think?

Sending this message here again because Bugzilla didn't know about the
mailing lists. Sorry for the noise!

===

Okay, here is some context about HostNotify. It is a rarely used SMBus
extension which allows clients to send a notification to the host. The
host must be able to listen to the I2C address 0x08. The only host
controller which has implemented this natively is the i801 because the
hardware sets a bit when this happened. (Sidenote, the only clients I am
aware of which send notifications are some touchscreen controllers.)
This is why the i801 driver calls i2c_handle_smbus_host_notify()
directly. And only that one, so far.

But recently, Alain Volmat got the idea that we can use the generic
slave framework to make host controllers listen on address 0x08 and
check for a valid HostNotification. This is why the generic
i2c_slave_host_notify_cb() calls i2c_handle_smbus_host_notify(), too.
This allows all I2C host controllers which select I2C_SLAVE to use
HostNotify. Those are few currently, but their number is steadily
increasing.

And as it looks to me, currently all drivers selecting I2C_SLAVE check
their interrupts which handle the slave state (i.e. managing I2C address
0x08) in a non-threaded context. But there is no guarantee for that.
Unless we formulate one. However, my gut feeling is that option #3 might
be not so much churn for this case, but I need to double check if I am
overlooking something.

Given that only some touchscreen controllers send HostNotify and you
need to enforce threaded irqs for this WARN, this might explain why it
went unnoticed for 10 years.

I hope this helps. Thank you everyone for the input provided so far!