iio: trigger: Fix a scheduling whilst atomic issue seen on tsc2046

IIO triggers are software IRQ chips that split an incoming IRQ into
separate IRQs routed to all devices using the trigger.
When all consumers are done then a trigger callback reenable() is
called.  There are a few circumstances under which this can happen
in atomic context.

1) A single user of the trigger that calls the iio_trigger_done()
function from interrupt context.
2) A race between disconnecting the last device from a trigger and
the trigger itself sucessfully being disabled.

To avoid a resulting scheduling whilst atomic, close this second corner
by using schedule_work() to ensure the reenable is not done in atomic
context.

Note that drivers must be careful to manage the interaction of
set_state() and reenable() callbacks to ensure appropriate reference
counting if they are relying on the same hardware controls.

Deliberately taking this the slow path rather than via a fixes tree
because the error has hard to hit and I would like it to soak for a while
before hitting a release kernel.

Signed-off-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Pengutronix Kernel Team <kernel@pengutronix.de>
Cc: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Tested-by: Oleksij Rempel <o.rempel@pengutronix.de>
Cc: <Stable@vger.kernel.org>
Link: https://lore.kernel.org/r/20211017172209.112387-1-jic23@kernel.org

1 files changed, 35 insertions, 1 deletions

diff --git a/drivers/iio/industrialio-trigger.c b/drivers/iio/industrialio-trigger.c
index b23caa2f2aa1..d3bdc9800b4a 100644
--- a/drivers/iio/industrialio-trigger.c
+++ b/drivers/iio/industrialio-trigger.c
@@ -162,6 +162,39 @@ static struct iio_trigger *iio_trigger_acquire_by_name(const char *name)
 	return trig;
 }
 
+static void iio_reenable_work_fn(struct work_struct *work)
+{
+	struct iio_trigger *trig = container_of(work, struct iio_trigger,
+						reenable_work);
+
+	/*
+	 * This 'might' occur after the trigger state is set to disabled -
+	 * in that case the driver should skip reenabling.
+	 */
+	trig->ops->reenable(trig);
+}
+
+/*
+ * In general, reenable callbacks may need to sleep and this path is
+ * not performance sensitive, so just queue up a work item
+ * to reneable the trigger for us.
+ *
+ * Races that can cause this.
+ * 1) A handler occurs entirely in interrupt context so the counter
+ *    the final decrement is still in this interrupt.
+ * 2) The trigger has been removed, but one last interrupt gets through.
+ *
+ * For (1) we must call reenable, but not in atomic context.
+ * For (2) it should be safe to call reenanble, if drivers never blindly
+ * reenable after state is off.
+ */
+static void iio_trigger_notify_done_atomic(struct iio_trigger *trig)
+{
+	if (atomic_dec_and_test(&trig->use_count) && trig->ops &&
+	    trig->ops->reenable)
+		schedule_work(&trig->reenable_work);
+}
+
 void iio_trigger_poll(struct iio_trigger *trig)
 {
 	int i;
@@ -173,7 +206,7 @@ void iio_trigger_poll(struct iio_trigger *trig)
 			if (trig->subirqs[i].enabled)
 				generic_handle_irq(trig->subirq_base + i);
 			else
-				iio_trigger_notify_done(trig);
+				iio_trigger_notify_done_atomic(trig);
 		}
 	}
 }
@@ -535,6 +568,7 @@ struct iio_trigger *viio_trigger_alloc(struct device *parent,
 	trig->dev.type = &iio_trig_type;
 	trig->dev.bus = &iio_bus_type;
 	device_initialize(&trig->dev);
+	INIT_WORK(&trig->reenable_work, iio_reenable_work_fn);
 
 	mutex_init(&trig->pool_lock);
 	trig->subirq_base = irq_alloc_descs(-1, 0,