Re: [alsa-devel] [PATCH v3 2/4] ASoc: kirkwood: merge kirkwood-i2c and kirkwood-dma
On Wed, Jul 31, 2013 at 08:18:06AM +0200, Jean-Francois Moine wrote:
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.
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.
The bias levels are stuck at off/standby all the time, which makes any kind of PM with spdif-dit impossible - or even any routing between stream input and output.
Basically, I'm tearing my hair out today looking at this stuff and getting nowhere fast - all I'm doing is finding more and more problems.
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.
ASoC: dapm: Treat DAI widgets like AIF widgets for power
seems to be the cause, this results in such setups having zero connected inputs/outputs reported, which causes them to remain powered down - because what used to happen before was the DAI links would report their powered state depending on whether they were active, and they are set active in soc_dapm_stream_event().
Now, when playback widget for the Codec and CPU DAIs get marked as active. The playback widget is created as a snd_soc_dapm_dai_in. It's power check function is set to dapm_dac_check_power. Since this widget is active, it checks for connected outputs via is_connected_output_ep().
We drop through to the second switch statement (why do we have two there? They're both switching on the same damned variable and its not like a widget can change its ID.) This is where the different behaviour has appeared - when these were just a simple snd_soc_dapm_dai, we used to just do the snd_soc_dapm_suspend_check() here, but we don't anymore. This is a snd_soc_dapm_dai_in type of widget, so we fall through this switch statement now and start searching the paths.
As these codecs have no paths, this ultimately ends up returning no connections. Hence why these codecs with no DAPM widgets declared but have PM support via the bias level stuff have stopped working.
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.
With that "fixed" (rather, altered to a state where it behaves the same way as it used to before the above commit) if I set up a DAPM route between the CPU DAI playback stream, an AIF (for spdif output) and the spdif-dit playback stream, I see the playback streams marked active and powered up, but the AIF stream remains powered down:
spdif-dit/dapm/Playback:Playback: On in 1 out 1 spdif-dit/dapm/Playback: stream Playback active spdif-dit/dapm/Playback: in "static" "spdif-playback" spdif-dit/dapm/bias_level:On
kirkwood-audio.1/dapm/spdif-playback:spdif-playback: Off in 1 out 0 kirkwood-audio.1/dapm/spdif-playback: in "static" "dma-playback" kirkwood-audio.1/dapm/spdif-playback: out "static" "Playback"
kirkwood-audio.1/dapm/dma-playback:dma-playback: On in 1 out 1 kirkwood-audio.1/dapm/dma-playback: stream dma-playback active kirkwood-audio.1/dapm/dma-playback: out "static" "spdif-playback" kirkwood-audio.1/dapm/dma-playback: out "static" "i2s-playback"
It looks to me like the DAPM stuff is - in one plain and simple word - buggered.
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.
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:
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.
participants (2)
-
Mark Brown -
Russell King - ARM Linux