On Tue, Jan 09, 2024 at 10:03:02PM +0000, James Ogletree wrote:

Hi Dmitry,

Thank you for your excellent review. Just a few questions.

quoted

On Jan 6, 2024, at 7:58 PM, Dmitry Torokhov [off-list ref] wrote:

On Thu, Jan 04, 2024 at 10:36:37PM +0000, James Ogletree wrote:

quoted

quoted

+
+ info->add_effect.u.periodic.custom_data = kcalloc(len, sizeof(s16), GFP_KERNEL);
+ if (!info->add_effect.u.periodic.custom_data)
+ return -ENOMEM;
+
+ if (copy_from_user(info->add_effect.u.periodic.custom_data,
+   effect->u.periodic.custom_data, sizeof(s16) * len)) {
+ info->add_error = -EFAULT;
+ goto out_free;
+ }
+
+ queue_work(info->vibe_wq, &info->add_work);
+ flush_work(&info->add_work);

I do not understand the need of scheduling a work here. You are
obviously in a sleeping context (otherwise you would not be able to
execute flush_work()) so you should be able to upload the effect right
here.

Scheduling work here is to ensure its ordering with “playback" worker
items, which themselves are called in atomic context and so need
deferred work. I think this explains why we need a workqueue as well,
but please correct me.

quoted

quoted

+
+static int vibra_playback(struct input_dev *dev, int effect_id, int val)
+{
+ struct vibra_info *info = input_get_drvdata(dev);
+
+ if (val > 0) {

value is supposed to signal how many times an effect should be repeated.
It looks like you are not handling this at all.

For playbacks, we mandate that the input_event value field is set to either 1

I am sorry, who is "we"?

or 0 to command either a start playback or stop playback respectively.
Values other than that should be rejected, so in the next version I will fix this
to explicitly check for 1 or 0.

No, please implement the API properly.

quoted

quoted

+ info->start_effect = &dev->ff->effects[effect_id];
+ queue_work(info->vibe_wq, &info->vibe_start_work);

The API allows playback of several effects at once, the way you have it
done here if multiple requests come at same time only one will be
handled.

I think I may need some clarification on this point. Why would concurrent
start/stop playback commands get dropped? It seems they would all be
added to the workqueue and executed eventually.

You only have one instance of vibe_start_work, as well as only one
"slot" to hold the effect you want to start. So if you issue 2 request
back to back to play effect 1 and 2 you are likely to end with
info->start_effect == 2 and that is what vibe_start_work handler will
observe, effectively dropping request to play effect 1 on the floor.

quoted

quoted

+ } else {
+ queue_work(info->vibe_wq, &info->vibe_stop_work);

Which effect are you stopping? All of them? You need to stop a
particular one.

Our implementation of “stop” stops all effects in flight which is intended.
That is probably unusual so I will add a comment here in the next
version.

Again, please implement the driver properly, not define your own
carveouts for the expected behavior.

Thanks.

-- 
Dmitry