On 8/3/20 8:44 AM, Christian Eggers wrote:
Hi Lars,
On Monday, 3 August 2020, 08:37:43 CEST, Lars-Peter Clausen wrote:
The sysfs IIO trigger uses irq_work to schedule the iio_trigger_poll()
and the promise of irq_work is that the callback will run in hard IRQ
context. That's the whole point of it.
irq_work_run_list(), which shows up in your callgraph, has as
BUG_ON(!irqs_disabled())[1], so we should never even get to calling
iio_trigger_poll() if IRQs where not disabled at this point. That's the
same condition that triggers the WARN_ON() in __handle_irq_event_percpu.
is my patch sufficient, or would you prefer a different solution?
The code in normal upstream is correct, there is no need to patch it
since iio_sysfs_trigger_work() always runs with IRQs disabled.
Are you using a non-upstream kernel? Maybe a RT kernel?
I use v5.4.<almost-latest>-rt
That explains it. Have a look at
0200-irqwork-push-most-work-into-softirq-context.patch.
The right fix for this issue is to add the following snippet to the RT
patchset.
diff --git a/drivers/iio/trigger/iio-trig-sysfs.c
b/drivers/iio/trigger/iio-trig-sysfs.c
--- a/drivers/iio/trigger/iio-trig-sysfs.c
+++ b/drivers/iio/trigger/iio-trig-sysfs.c
@@ -161,6 +161,7 @@ static int iio_sysfs_trigger_probe(int id)
iio_trigger_set_drvdata(t->trig, t);
init_irq_work(&t->work, iio_sysfs_trigger_work);
+ t->work.flags = IRQ_WORK_HARD_IRQ;
ret = iio_trigger_register(t->trig);
if (ret)
