+alsa-devel list

On 04/25/2012 01:19 AM, Tomi Valkeinen wrote:

On Tue, 2012-04-24 at 23:48 -0500, Ricardo Neri wrote:

...

On 04/23/2012 08:01 AM, Tomi Valkeinen wrote:

...

On Wed, 2012-03-28 at 16:38 -0600, Ricardo Neri wrote:

...

Implement the DSS device driver audio support interface in the HDMI panel driver and generic driver. The implementation relies on the IP-specific functions that are defined at DSS probe time.

A HW-safe spinlock is used to protect the audio functions. This is because

What is a "HW-safe spinlock"?

Sorry, I meant a spinlock that disables the HW irqs when held:hardirq-safe.

...

...

the audio functions may be called while holding a lock; in such case, the panel's driver mutex is not suitable. Functions should be used to set registers and should not wait for any other event.

Are you sure this is the only option? What lock is being held?

For instance, ALSA calls the start audio function while holding a hardirq-safe readlock. Hence, when reaching the HDMI panel start function, a lock is held and irqs are disabled. Using a mutex, that might sleep, is not correct; nor it is using an hardirq-unsafe spinlock. Otherwise, deadlocks and/or inverse lock ordering may arise. This situation was signaled by lockdep.

IMHO, as the DSS device driver does not know who is going to use it (at least the audio part), it should not assume that no locks are held when its functions are called.

...

While a spinlock may be ok for now, quite often enabling/disabling things do not happen immediately,and it's much easier to do the wait synchronously.

I don't understand this comment. To me, holding a lock until the enabling function returns is synchronous. Would you please clarify?

I meant that quite often when enabling things on hardware it takes time until the HW is actually up and running. Perhaps a regulator needs to be started, or such. And it's usually simpler to wait for the HW to be up synchronously in the enable function, instead of some kind of asynchronous mechanism. And if a function waits synchronously, a mutex is better than spinlock.

And in that sense it's often better to define (define meaning, adding a comment, or just mentally taking note about it) that the functions in the API may sleep, so that the driver is free to do what is best for the case, and it's also future-proof in a way that you can easily add function calls that sleep to the functions in the future.

But you are also right in your previous comment, it's better to define functions so that they never sleep, as that gives the callers the freedom to call the funcs in atomic context.

Perhaps we can have audio_start() that never sleeps, it just enables a few bits that start the audio. But how about audio_enable()? Is it possible that on some OMAP version it would need to enable a regulator, or set a gpio that's in an external i2c controlled mux chip, or such?

I think it might be possible to have such a scenario if, for instance, an external chip is used to support DisplayPort on OMAP4/5. As DisplayPort can support audio-only use-cases, it would have to enable the adapter chip (but HDMI output would have to be enabled to feed the chip, though).

If so, we need to make sure it's not currently called in an atomic context, because it would break if the function will sleep in the future. And with "make sure" I just mean that we check the code and keep it in mind. Or perhaps adding a comment in the header, that says "audio_enable may sleep, other audio functions do not sleep" or such.

I revisited the ALSA code. Only the audio_start function is atomic. Although ALSA may not be the only user, it makes sense to me to think that they will follow a similar approach in terms of locks.

Hence, based on that and on the reasons you describe (audio_enable potentially taking too long to return), Rephrasing what you stated, a mutex may be used for the enable/disable and config operations. Only start/stop would be protected by a spinlock. This should be described in comments in the header file. Does it make sense to you?

BR,

Ricardo

Tomi