Note that CEC on Linux for some reason doesn't work after resume. I'll look into that. However, all important registers are preserved.

I am okay with depending on Linux to initialize the hardware, as long as the driver is disabled by default, and that dependency is noted in the Kconfig help text.

Actually, apart from clocks and resets (in CCU and additional ones in controller), there is nothing to initialize. Since I'm running this with debug uart and IR enabled, I only worry that in deeper sleep modes this won't work. Bus must still be responsive, which means 24 MHz clock must be running.

Before merging, I would like to resolve (or at least understand) the issue preventing CEC from working after resume.

Sure, me too.

Code should be much more readable now with magic numbers moved to macros. Probably not all wake up messages should be enabled. For one or two doesn't really make sense IMO. Linux issue was on my side. I removed one patch, which I thought it would be needed, but it was not. CEC now works correctly after resume.

All in all, this works pretty well, except when debug uart is disabled (no wake up source works, not even the button). I guess driver should make sure that 24 MHz clock is never stopped. Any ideas how to correctly do that? I speculate that H6 has some changes in this department to allow deeper sleep modes and it might need different handling than A64, H3 and H5.

I suppose defining HDMI_DDC_CLK_GATING clock (offset 0x154 in A64 CCU) and assigning it 24 MHz clock as parent would be enough.

CEC now works correctly after resume.

🎉

Bus must still be responsive, which means 24 MHz clock must be running.

All of the buses are reparented to OSC32K, so MMIO register access should generally work after OSC24M is stopped.

All in all, this works pretty well, except when debug uart is disabled (no wake up source works, not even the button).

Hmm, this is a bit surprising.

Since you mention you have an Orange Pi Win: you can select CONFIG_SERIAL_DEV_R_UART and use the pins on the 40-pin header to get a debug console that works at any suspend depth. Of course, since R_UART runs off IOSC, you will have to lower the baud rate. Either 38400 or 9600 works for me, depending on the board.

I guess driver should make sure that 24 MHz clock is never stopped. Any ideas how to correctly do that?

You can limit the suspend depth by taking a reference to CLK_OSC24M in the driver. See select_suspend_depth() in common/system.c. Or you can change that function if needed.

I suppose defining HDMI_DDC_CLK_GATING clock (offset 0x154 in A64 CCU) and assigning it 24 MHz clock as parent would be enough.

This would more accurately describe the hardware, but it would create other issues, as Crust would gate the clock when it gets released, thus breaking Linux. So I suggest using CLK_OSC24M directly.

I speculate that H6 has some changes in this department to allow deeper sleep modes and it might need different handling than A64, H3 and H5.

Yes, H6 has a separate HDMI_CEC_CLK that runs off OSC32K. So once the driver has a clock consumer, taking advantage of that should be easy.

All of the buses are reparented to OSC32K, so MMIO register access should generally work after OSC24M is stopped.

Documentation that I have states that isfrclk clock must be between 18 MHz and 27 MHz. While it's not described anywhere how AW integrated this core, I suspect that isfrclk means input for special function registers clock and it would be logical that it correlates with HDMI_DDC_CLK_GATING.

There is another possible thing that's missing in DT. DW HDMI controller binding mentions another optional clock - CEC, which is not specified in DTs for 40 nm SoCs. Documentation states it must be connected to 32768 Hz source. This is of course 32k clock from RTC. I found that when searching for cause why CEC doesn't work on some H3 boards and does on others. It correlates with board having external 32k crystal or not. I also forced internal RC oscillator on board with external crystal and CEC stopped working soon after. Also adjusting RC oscillator by hand also caused CEC to work again. Anyway, that's good enough reason to update all those SoC DTSI files with additional CEC clock, connected to 32k clock.

Yes, H6 has a separate HDMI_CEC_CLK that runs off OSC32K. So once the driver has a clock consumer, taking advantage of that should be easy.

If you check Linux H6 CCU driver, you'll notice that I forced CEC clock to PLL periph0 x2, because 32k didn't work (BSP Linux driver does the same). Additionally, boards like Tanix TX6 don't have external 32k crystal at all and 32k crystal on Beelink GS1 doesn't work always, so it's disabled.

This would more accurately describe the hardware, but it would create other issues, as Crust would gate the clock when it gets released, thus breaking Linux. So I suggest using CLK_OSC24M directly.

Noted.

Claiming and releasing CLK_OSC24M works. So this code seems good for A64. Testing on other SoCs, especially on H6 is still needed. Probably next week...

I tested this on Tanix TX6 (H6) and it works too.

@jernejsk Good to hear! Do you think this is ready to merge?

I can fix up the CI failure (minor style checks) when applying.

Yeah, I think automatic mode is as good as it can get, so this PR is ready to be merged. There are downsides like waking up even if some other device was addressed, but such issues can only be solved by manual mode. I'll work on it after I get a board for sniffing HDMI signals.

BTW, do you think is possible to continuously doing RC oscillator calibration in standby, like every 5 min or so? This would allow using CEC even on boards without external 32k crystal. CEC standard allows for about 5% deviation but RC oscillator is often worse.

Sounds like a good plan to me! I've merged this as 2e5f355.

Yes, periodic calibration is a good idea. I wrote up #198 to cover that, since I'm closing this PR.