On Mon, Feb 13, 2012 at 02:56:55PM +0100, Takashi Iwai wrote:
Mark Brown wrote:
There's rather a lot of people who do embedded development and do need to look at this stuff, though. There's a real cost to raising the bar for humans to understand things.
Hm, I'm afraid that this is getting apart from the real worth discussion. I'm interested in whether "Line-Out" is acceptable enough, or it's a matter of taste, just like Marmite is acceptable as the normal English breakfasts.
Well, I dunno - you did suggest adding direction modifiers to the jack names which would both be a more general solution to the problem and clean up the naming issue.
with the input-jack layer. The input-jack doesn't provide the index number. So, you can't register the same name correctly. Thus I tried to create a unique name string as input-jack id.
It's really never been the intention to have an "input-jack" layer - the API is there to hide all the input layer stuff it can from the individual audio drivers. The only thing that leaks out is the button remapping which I couldn't think of anything more sensible to do with them.
Now, looking back to your implementation, the kctl jack is created with a name string based on the switch type, not from the name string passed to snd_jack_new() itself. Why not passing the string as is?
We can't pass it as-is since we'll need multiple controls per jack - literally all the systems I have available to test on would fail if they used the same name as they all have headset jacks. As I said further up the thread my plan had been to refactor the name generation code once the separate kctl code is gone and they can be merged into the same source file which makes this a lot easier.
With these patches I'm mainly looking for a quick and simple refactoring to get rid of the duplication - once we've got rid of the duplicate driver APIs then we can refactor the internals of the implementation however we like but there's a pretty pressing need to fix the driver API and get all the drivers doing the same thing to userspace even if it's not ideal.
If multiple kctls are needed (e.g. for headset), you can either create a new name or just create the multiple instances with the same name but give extra TLVs to identify which type it is. Or, you can create an enum control so that both the input-jack and the kctl-jack correspond as 1:1.
This stuff is why I've been asking you about the ABI and general semantics for the new controls, I was never sure what the intention was - when you said they were just plain booleans I just went ahead and created multiple controls as that seemed to be what the current idiom was.
I don't think an enumeration is going to scale so well given the number of possible combinations you can get on more complex jacks.
In other words, if 1:1 mapping isn't needed, we don't have to create a different name string at all. Just create "Jack Detect" controls with different indices and with corresponding TLVs. This would be enough (could be even easier) for a machine parser.
That'd work I think, though I would create a new control for each physical jack at least just for clarity.
And, it seems that there is no standard rule defined for the id string of input-jack. If 1:1 model is used, we'd need to define this rule.
The input stack doesn't have the same concept of machine parsable names for things that ALSA does (honestly that's pretty unique to ALSA within the kernel) - the names it users are all free form text intended to be meaningful to humans only.
BTW, I found another issue. With this implementation, you can't build kjack ctl without input-jack. I can work around it for HD-audio by adding more ifdef, but it'd be ugly...
I really don't see this as a problem; it seems vanishingly unlikely that any realistic system would want to disable the input API and need accessory detection. Obviously the end goal here is that you wouldn't be able to have HDA specific hacks in the code as everything will be handled by the ALSA core code...
To my mind we should just support everything and let userspace worry about deciding what to use, this isn't the sort of thing which should require a kernel rebuild to configure and it's not like there's any fundamental problem or massive overhead caused by having the devices show up.