On Sat, Aug 03, 2013 at 11:23:33PM +0100, Russell King - ARM Linux wrote:

On Wed, Jul 31, 2013 at 08:18:06AM +0200, Jean-Francois Moine wrote:

quoted

To avoid the declaration of a 'kirkwood-pcm-audio' device in the DT,
this patch merges the kirkwood-i2c and kirkwood-dma drivers into one
module associated with 'kirkwood-i2s'. 

I suggest holding off on this stuff at the moment.  I think Mark and Liam
(who now have a whole raft of emails from me today) have some work to do
to fix ASoC to work how they're saying it should work, because to me
ASoC seems to contradict what they're telling me.  To put it another way,
it must be buggy.

Or push this stuff in there and help fix any issues that come up, what
issues there are with the CPU side stuff are there mostly because nobody
has been upstreaming anything and hence exercising the code.  This merge
is going to be needed for the DT conversion anyway so I can't see any
gain in waiting, other things depend on this rather than the other way
around.

The DAPM stuff seems to be the worst thing since mouldy bread.  I'm
chasing what seems to be multiple bugs through this stuff, and many of
them are not particularly nice.

For example - we register multiple copies of DAPM widgets for the same
set of declarations if both a CPU DAI and a platform share the same
struct device, and thereby end up overwriting some pointers in the DAI
structure.  It seems that CPU DAIs themselves aren't supposed to have
DAPM widgets, but the core creates some... but there's no explicit
cleanup of them unlike the other DAPMs, and there's no debugfs for them.

CPU DAIs are supposed to be able to have widgets, I'm not sure why you
say they aren't.  It's possible this doesn't work since most hardware is
structured such that the control belongs in a DSP which tends to end up
in the platform device so even the out of tree users aren't likely to
exercise this path but from a framework point of view that seems like a
thing I'd expect to work and if it doesn't work we should make that so.

Removal paths are, by and by large, not really exercised since nobody
actually does that in production and most people's development model is
based on rebooting full kernels.

For example, where a codec has no input/output widgets declared, it used
to be powered up automatically by ASoC as a matter of course.  Things
like UDA134x and other things.  Things like spdif-dit.  That "mysteriously"
stopped happening.

This is DAPMless CODECs, frankly they're more trouble than they're
worth given how trivial it is to implement DAPM and the fact that they
constantly cause problems due to the special casing they need and the
fact that development is pretty much all happening on modern hardware
with CODEC drivers that support DAPM.  This isn't the first time they've
been missed, it won't be the last either.

Now, about the spdif-dit, if we're going to have to add "pin" widgets
to it, what the output of a SPDIF in terms of DAPM widgets?  At a guess,
it's a "codec pin" despite there being no codec and no "pin" on that
codec in reality, and that "pin" is always active.

There's an output from the S/PDIF device - the S/PDIF signal it sends,
either electrically or optically (usually the former).

It looks to me like the DAPM stuff is - in one plain and simple word -
buggered.

It does seem to be working adequately in a rather large number of
systems.  I think you're confusing DAPM with support for non-trivial
SoCs, currently all the mainline DAPM use is in CODEC devices.

As far as I can tell your problems here come from a combination of using
very old devices like the uda134x which haven't ever been particularly
well maintained and trying to do new things with the frameworks.  If
you're working on something that's not been done before it's likely that
the frameworks will need extending.

I've no idea what the right fixes are in this area.  It needs someone
like Mark or Liam who supposedly understand this to spend time checking
that it actually operates as they _think_ it should operate, because
at the moment it plainly doesn't.

This is all working fine for me with the systems I have available to me
in mainline - I don't have any systems at all with non-DAPM devices and
obviously nothing in mainline uses DPCM.