On 2022-03-16 13:01, Dmitry Osipenko wrote:
On 3/16/22 12:12, Sascha Hauer wrote:
On Mon, Mar 14, 2022 at 08:54:27PM +0300, Dmitry Osipenko wrote:
On 3/14/22 11:18, Sascha Hauer wrote:
On Sun, Mar 13, 2022 at 12:07:56AM +0300, Dmitry Osipenko wrote:
On 3/11/22 11:33, Sascha Hauer wrote:
The rk3568 HDMI has an additional clock that needs to be enabled for the HDMI controller to work. This clock is not needed for the HDMI controller itself, but to make the SoC internal bus logic work. From the reference manual:
> 2.8.6 NIU Clock gating reliance > > A part of niu clocks have a dependence on another niu clock in order to > sharing the internal bus. When these clocks are in use, another niu > clock must be opened, and cannot be gated. These clocks and the special > clock on which they are relied are as following: > > Clocks which have dependency The clock which can not be gated > ----------------------------------------------------------------- > ... > pclk_vo_niu, hclk_vo_s_niu hclk_vo_niu > ... The clock framework does not support turning on a clock whenever another clock is turned on, so this patch adds support for the dependent clock to the HDMI driver. We call it "NIU", which is for "Native Interface Unit"
This still doesn't make sense to me. You're saying that "pclk_vo_niu, hclk_vo_s_niu" depend on "hclk_vo_niu", but HDMI doesn't use pclk_vo, it uses pclk_hdmi.
pclk_hdmi_host is a child clock of pclk_vo:
aclk_vo 2 2 0 300000000 0 0 50000 Y aclk_hdcp 0 0 0 300000000 0 0 50000 N pclk_vo 2 3 0 75000000 0 0 50000 Y pclk_edp_ctrl 0 0 0 75000000 0 0 50000 N pclk_dsitx_1 0 0 0 75000000 0 0 50000 N pclk_dsitx_0 1 2 0 75000000 0 0 50000 Y pclk_hdmi_host 1 2 0 75000000 0 0 50000 Y pclk_hdcp 0 0 0 75000000 0 0 50000 N hclk_vo 2 5 0 150000000 0 0 50000 Y hclk_hdcp 0 0 0 150000000 0 0 50000 N hclk_vop 0 2 0 150000000 0 0 50000 NIt was unclear that the pclk_hdmi is the child of pclk_vo by looking at the clk driver's code, thank you!
Won't be better if the implicit clk dependency would be handled internally by the RK clk driver?
I have considered handling something like coupled clocks in the clock framework, but I have not yet considered handling this internally in the Rockchip clock driver.
I just had a quick look at the driver. While it sounds like an easy task to enable gate A whenever gate B is enabled I haven't found a good way to integrate that into the clk driver. It seems to me that it's not easier to implement in the clk driver than it is to implement it in the clk framework where it could be used by others as well.
For example, by making the common gate shared/refcounted. Have you considered this variant? Then we won't need to change the DT bindings.
For the DT bindings it is just an additional clock. Should we have a better way to handle that case in the future we could simply ignore the additional clock. I wouldn't bother too much about this.
To me that NIU quirk should be internal to the clk h/w module, so it doesn't feel nice to mix the clk h/w description with the HDMI h/w description.
On the other hand, making clk driver to handle this case indeed will take some effort as I see now. For example, clk driver of NVIDIA Tegra has concept of shared gates, but bringing it to the RK clk driver will be quite messy.
From a quick look, it seems like it could be straightforward conceptually at least. Presumably: subclass clk_gate_ops to enable/disable a required clock before enabling/disabling normally, have rockchip_clk_register_branch() resolve an optional required clock and pick gate_ops as appropriate, then the rest is basically just boilerplate for describing the dependencies in the first place. However I'd agree that in practical implementation terms it does look even simpler and cleaner for the clk_hw abstraction to provide the appropriate ops and resolution itself.
Alright, let's work around the clk limitation like you're suggesting. I agree that it shouldn't really be a problem to deprecate the extra clock later on.
If there's a realistic chance that someone will actually work on a proper coupled/dependent/whatever clock abstraction before the rest of RK3588 is supported well enough for mainline users to start really caring about power efficiency, then arguably the simplest and cleanest workaround would be the other option that Elaine mentioned, of just marking hclk_vo as critical for now. If it's likely to turn into a "nothing's as permanent as a temporary fix" situation, though, then the DT binding has less functional impact, even if it does leave us developers with baggage down the line.
Robin.
I have last questions..
- Previously you said that the PD driver takes care of enabling all the
clocks it can find in the device by itself on RPM-resume, then why HDMI driver needs to enable the clock explicitly?
- You added clk_prepare_enable(), but there is no corresponding
clk_disable_unprepare(), AFAICS. Why?