On Wed, 27 Jun 2018 12:57:22 +0200, Sriram Periyasamy wrote:
On Wed, Jun 27, 2018 at 11:34:28AM +0200, Takashi Iwai wrote:
On Wed, 27 Jun 2018 11:17:55 +0200, Sriram Periyasamy wrote:
On Wed, Jun 27, 2018 at 10:48:05AM +0200, Takashi Iwai wrote:
On Wed, 27 Jun 2018 10:35:35 +0200, Sriram Periyasamy wrote:
On Tue, Jun 26, 2018 at 01:33:36PM +0200, Takashi Iwai wrote:
On Tue, 26 Jun 2018 13:24:01 +0200, Sriram Periyasamy wrote: > > On Mon, Jun 25, 2018 at 05:18:27PM +0200, Takashi Iwai wrote: > > On Mon, 25 Jun 2018 16:51:12 +0200, > > Sriram Periyasamy wrote: > > > > > > On Mon, Jun 25, 2018 at 11:53:59AM +0200, Takashi Iwai wrote: > > > > On Mon, 25 Jun 2018 11:23:17 +0200, > > > > Sriram Periyasamy wrote: > > > > > > > > > > On Mon, Jun 25, 2018 at 11:08:13AM +0200, Takashi Iwai wrote: > > > > > > On Mon, 25 Jun 2018 10:53:49 +0200, > > > > > > Sriram Periyasamy wrote: > > > > > > > > > > > > > > On Fri, Jun 22, 2018 at 07:14:50PM +0200, Takashi Iwai wrote: > > > > > > > > On Fri, 22 Jun 2018 18:42:24 +0200, > > > > > > > > Pierre-Louis Bossart wrote: > > > > > > > > > > > > > > > > > > On 6/22/18 4:25 AM, Sriram Periyasamy wrote: > > > > > > > > > > On certain platforms, Display HDMI HDA codec was not going to sleep state > > > > > > > > > > after the use when links are powered down after turning off the display > > > > > > > > > > power. As per the HW recommendation, links are powered down before turning > > > > > > > > > > off the display power to ensure that the codec goes to sleep state. > > > > > > > > > > > > > > > > > > The change looks ok, I am still not clear on display power > > > > > > > > > management. snd_hdac_display_power is called both from hdac_hdmi.c and > > > > > > > > > skl.c in the respective probe/suspend/resume routines, wondering if > > > > > > > > > this is redundant or intentional. > > > > > > > > > > > > > > > > Yeah, it looks like that skl.c keeps the display power on by assuming > > > > > > > > that it's turned off by the codec driver, but it doesn't happen > > > > > > > > actually. I also find it fishy there... > > > > > > > > > > > > > > > > > > > > > > snd_hdac_display_power is called intentionally to ensure that the codec is > > > > > > > powered up before any communication is initiated from either skl.c or > > > > > > > hdac_hdmi.c. > > > > > > > > > > > > But who turns it down if the device is not used? > > > > > > > > > > Powering down of codec is taken care in codec's suspend routine. > > > > > > > > But you'll power up twice, no? Once in skl_resume() (which is no > > > > runtime resume but only for S3/S4) and once in codec resume. > > > > I wonder who will turn down when you wake up from S3. > > > > > > No. When entering to S3, hdmi_codec_prepare() will invoke the runtime > > > resume of codec hdac_hdmi_runtime_resume() which power up the codec and > > > skl_suspend power down the codec. When it wakes up from S3, skl_resume() > > > powers up the codec and hdmi_codec_complete invokes the runtime suspend of > > > codec hdac_hdmi_runtime_suspend() which powers down the codec. > > > > Hm, but what about the i915 power well refcount, not the codec runtime > > PM refcount? Does the refcount reach to zero when the system goes to > > runtime suspend? That puzzles me. > > > > Yes, i915 power well refcount goes to zero when system goes to runtime > suspend. The value of bus->i915_power_refcount in snd_hdac_display_power > is zero. > > > > This design is required to achieve the runtime power reference count for > > > codec device when it is put to S3/S4 while playback was going on. > > > > ... and this "design" picture is completely missing. > > That's why illustrating it was requested instead of a patch :) > > > > > > The runtime > > > > suspend doesn't call snd_hdac_display_power() in skl.c. > > > > > > > > > > Display power management for all runtime uses cases are handled in the > > > codec driver hdac_hdmi.c. > > > > But it means that hdac_hdmi.c does manage the extra decrement the > > refcount that was taken in skl.c? > > > > No, snd_hdac_display_power is invoked in skl.c for the following cases, > > skl_probe_work - display power on, enumurate the codec and display power > off > skl_suspend & skl_resume - display power off/on for S3/S0 cycle > respectively > > snd_hdac_display_power is invoked in hdac_hdmi.c for the following cases, > > hdac_hdmi_dev_probe - display power on, probe the device and put the > device to runtime suspend explicitly > > runtime suspend & resume - display power off and on respectively
That puzzled me. How does display refcount goes to zero at runtime suspend of skl? skl_suspend() is only for S3, not for runtime PM.
Runtime suspend and resume of skylake doesn’t perform reference counting of display. For an HDMI playback stream, below is the sequence.
Open:
- skylake runtime resume
- hdac_hdmi runtime resume --> ref count is incremented
Close:
- hdac_hdmi runtime suspend --> ref count is decremented
- skylake runtime suspend
Suppose you probe skl, hdac_hdmi, then do runtime suspend hdac_hdmi and skl respectively. How does the display pm refcount go to zero?
In skylake probe, the display ref count is incremented at start of probe and decremented at end of probe. The ref count decrement is not performed in skylake runtime suspend. In hdac_hdmi probe, the display refcount is incremented in probe, and decremented in hdac_hdmi runtime suspend.
Well, the display refcount is incremented in S3 resume in skl.c. Who decrements it at runtime suspend?
It is decremented at runtime suspend of codec hdac_hdmi_runtime_suspend(). Below are sequence,
S3 cycle without playback going on,
Entering S3,
- hdmi_codec_prepare() invoke the runtime resume of codec
- hdac_hdmi_runtime_resume() --> increments the refcount
- skl_suspend() -- > decrements the refcount
Wake up from S3,
- skl_resume() --> increments the refcount
- hdmi_codec_complete() invokes the runtime suspend of codec
- hdac_hdmi_runtime_suspend() -- > decrements the refcount
So you're calling the runtime suspend even though it's not runtime resumed? That confuses readers and needs a clear explanation somewhere in the code.
Yes, we will add a separate patch for this.
That'd be helpful.
And what's the case if multiple codecs are present?
We need to check on multiple codecs scenario. Do you think any issues upfront?
No, and maybe my question was moot - it's only a single GPU, so always a single codec, supposedly.
Takashi