[alsa-devel] [PATCH] ASoC: intel: cht_bsw_max98090_ti: Enable codec clock once and keep it enabled

[Hans de Goede]

Hi,

On 24-01-19 16:02, Pierre-Louis Bossart wrote:
> 
> On 1/24/19 4:34 AM, Hans de Goede wrote:
>> Users have been seeing sound stability issues with max98090 codecs since:
>> commit 648e921888ad ("clk: x86: Stop marking clocks as CLK_IS_CRITICAL")
>>
>> At first that commit broke sound for Chromebook Swanky and Clapper models,
>> the problem was that the machine-driver has been controlling the wrong
>> clock on those models since support for them was added. This was hidden by
>> clk-pmc-atom.c keeping the actual clk on unconditionally.
>>
>> With the machine-driver controlling the proper clock, sound works again
>> but we are seeing bug reports describing it as: low volume,
>> "sounds like played at 10x speed" and instable.
>>
>> When these issues are hit the following message is seen in dmesg:
>> "max98090 i2c-193C9890:00: PLL unlocked".
>>
>> Attempts have been made to fix this by inserting a delay between enabling
>> the clk and enabling and checking the pll, but this has not helped.
>>
>> It seems that if we ever disable the clk, the pll looses its lock and
>> then we get various issues.
>>
>> This commit fixes this by enabling the clock once at probe time,
>> in essence restoring the old behavior of clk-pmc-atom.c always keeping
>> the clk on, but then only on machines with a max98090 codec.
>>
>> Fixes: 648e921888ad ("clk: x86: Stop marking clocks as CLK_IS_CRITICAL")
>> Reported-and-tested-by: Mogens Jensen <mogens-jensen at protonmail.com>
>> Reported-and-tested-by: Dean Wallace <duffydack73 at gmail.com>
>> Signed-off-by: Hans de Goede <hdegoede at redhat.com>
> 
> Sorry, I am not really convinced this patch can be applied as is.
> 
> 1. All these clock changes were introduced to solve S0i1 issues, so now if we keep the clocks on isn't this going to impact suspend/resume usages?
> Has anyone looked into this?

Yes this will cause the affected devices to not sleep deeper then S0i1, just as
they did not sleep deeper before the clock changes. This seems to be the
lesser of the 2 evils. This patch is about fixing the sound regressions
caused by the clock changes on devices with the max98090 codec. It seems
that going back to the old situation where the clock is permanently
enabled is the best way to fix the regression.

> I am also wondering what the assumptions were on what the firmware does, and if the issues occur with the legacy firmware or Coreboot?

AFAIK both reporters are using coreboot, but this seems unrelated to the
firmware, the problem is the PLL loosing lock when we turn off the clock
and not re-locking when we turn it back on.

> 2. Also the fix is incorrect in that the clock needs to be set to 19.2 MHz for Baytrail devices (Rambi and friends) which reuse this machine driver. Even if there was agreement on leaving the clocks on, the clk_set_rate() call needs to be added back (default is 25 MHz on Baytrail).

I was assuming the clock would be setup correctly by the firmware since
at least on the 2 devices for which I've been getting reports we were
setting the clk-rate of the wrong clock, so in essence we were doing
a no-op.

But if you're worried about other devices then I can do a v2 re-adding
the setting of the clk-rate. We can simply move the old-code to do
that to the same point where we do the one-time enable now.

According to the 2 users I've been talking to, there have been a lot
of issues with the PLL-lock problem, so this likely affects all platforms.
Alternatively we could only do the enable-once thing for the platforms
with the QUIRK_PMC_PLT_CLK_0 quirk, as those are the ones for which
we've been receiving bug reports.

Regards,

Hans