On 2/11/21 5:56 PM, William Breathitt Gray wrote:
On Wed, Dec 30, 2020 at 11:36:45AM -0600, David Lechner wrote:
On 12/25/20 6:15 PM, William Breathitt Gray wrote:
diff --git a/Documentation/ABI/testing/sysfs-bus-counter-104-quad-8
b/Documentation/ABI/testing/sysfs-bus-counter-104-quad-8
index eac32180c40d..0ecba24d43aa 100644
--- a/Documentation/ABI/testing/sysfs-bus-counter-104-quad-8
+++ b/Documentation/ABI/testing/sysfs-bus-counter-104-quad-8
@@ -1,3 +1,28 @@
+What: /sys/bus/counter/devices/counterX/countY/irq_trigger
Do we really need this sysfs attribute? Shouldn't interrupts be configured
_only_ by the chrdev interface?
I think this attribute can go away because we can implicitly figure out
the correct IRQ configuration from the struct counter_watch data when a
user executes a COUNTER_ADD_WATCH_IOCTL ioctl command.
However, I need some help deciding on an appropriate behavior for
conflicting counter_watch configurations. Let me give some context
first.
The 104-QUAD-8 features 8 channels (essentially 8 independent physical
counters on the device). Each channel can independently issue an event,
but any particular channel can only be set to a single kind of event
(COUNTER_EVENT_INDEX, COUNTER_EVENT_OVERFLOW, etc.).
The purpose of the irq_trigger sysfs attribute I introduced in this
patch is to allow the user to select the event configuration they want
for a particular channel. We can theoretically figure this out
implicitly from the struct counter_watch request, so this sysfs
attribute may not be necessary.
However, how do we handle the case where a user executes two
COUNTER_ADD_WATCH_IOCTL ioctl commands for the same channel but with
different event selections? I'm considering three possible behaviors:
* Fail the second ioctl call; event selection of the first struct
counter_watch takes precedence and thus second is incompatible.
* Issue a dev_warn() indicating that the second struct counter_watch
event selection will now be the event configuration for that channel.
* Don't notify the user, just silently reconfigure for the second struct
counter_watch event selection.
I'm suspecting the first behavior I listed here (ioctl returning failed)
is the most appropriate as a user is explicitly made known of this
particular device's inability to support more than one type of event per
channel.
What do you think?
I agree that it should return an error instead of adding the watch.
I'm pretty sure that is how I implemented the TI eQEP driver already.
