On Sun, Jun 20, 2021 at 11:04:23PM +0200, Wolfram Sang wrote:

On Fri, Jun 04, 2021 at 04:27:44PM -0700, Stephen Boyd wrote:

quoted

From: Dmitry Torokhov <dmitry.torokhov@gmail.com>

If an i2c client receives an interrupt during reboot or shutdown it may
be too late to service it by making an i2c transaction on the bus
because the i2c controller has already been shutdown. This can lead to
system hangs if the i2c controller tries to make a transfer that is
doomed to fail because the access to the i2c pins is already shut down,
or an iommu translation has been torn down so i2c controller register
access doesn't work.

Let's simply disable the irq if there isn't a shutdown callback for an
i2c client when there is an irq associated with the device. This will
make sure that irqs don't come in later than the time that we can handle
it. We don't do this if the i2c client device already has a shutdown
callback because presumably they're doing the right thing and quieting
the device so irqs don't come in after the shutdown callback returns.

Cc: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Reported-by: kernel test robot <redacted>
[swboyd@chromium.org: Dropped newline, added commit text, added
interrupt.h for robot build error]
Signed-off-by: Stephen Boyd <redacted>

I think this is for-current material because it fixes hangs. Or better
for-next to check for side-effects?

We lived with this issue for many years, so letting it cook in next will
not hurt us IMO.

Thanks.

-- 
Dmitry