Re: [PATCH v6] clk: qcom: Add lpass clock controller driver for SDM845

[Stephen Boyd]

Quoting Taniya Das (2018-10-17 05:04:10)
> 
> 
> On 10/17/2018 5:07 PM, Taniya Das wrote:
> > Hello Stephen,
> > 
> > On 10/12/2018 11:05 PM, Stephen Boyd wrote:
> >> Quoting Taniya Das (2018-10-09 23:12:27)
> >>>
> >>>
> >>> On 10/10/2018 2:22 AM, Stephen Boyd wrote:
> >>>> Quoting Taniya Das (2018-10-09 10:26:38)
> >>>>> Hello Stephen,
> >>>>>
> >>>>> On 10/8/2018 8:14 AM, Stephen Boyd wrote:
> >>>>>> Quoting Taniya Das (2018-10-04 05:02:26)
> >>>>>>> Add support for the lpass clock controller found on SDM845 based 
> >>>>>>> devices.
> >>>>>>> This would allow lpass peripheral loader drivers to control the 
> >>>>>>> clocks to
> >>>>>>> bring the subsystem out of reset.
> >>>>>>> LPASS clocks present on the global clock controller would be 
> >>>>>>> registered
> >>>>>>> with the clock framework based on the device tree flag. Also do 
> >>>>>>> not gate
> >>>>>>> these clocks if they are left unused.
> >>>>>>
> >>>>>> Why not gate them? This statement states what the code is doing, 
> >>>>>> not why
> >>>>>> it's doing it which is the more crucial information that should be
> >>>>>> described in the commit text. Also, please add a comment about it 
> >>>>>> to the
> >>>>>> code next to the flag.
> >>>>>>
> >>>>>> I am concerned that it doesn't make any sense though, so probably it
> >>>>>> shouldn't be marked as CLK_IGNORE_UNUSED and it's papering over some
> >>>>>> other larger bug that needs to be fixed.
> >>>>>>
> >>>>>
> >>>>> It does not have any bug, it is just that to access these lpass
> >>>>> registers we would need the GCC lpass registers to be enabled. I would
> >>>>> update the same in the commit text.
> >>>>>
> >>>>> During clock late_init these clocks should not be accessed to check 
> >>>>> the
> >>>>> clock status as they would result in unclocked access. The client 
> >>>>> would
> >>>>> request these clocks in the correct order and it would not have any 
> >>>>> issue.
> >>>>>
> >>>>
> >>>> That seems like the bug right there. If the LPASS registers can't be
> >>>> accessed unless the clks in GCC are enabled then this driver needs to
> >>>> turn the clks on before reading/writing registers. Marking the clks as
> >>>> ignore unused is skipping around the real problem.
> >>>>
> >>>
> >>> If the driver requests for the clocks they would maintain the order. But
> >>> if the clock late init call is invoked before the driver requests, there
> >>> is no way I could manage this dependency, that is the only reason to
> >>> mark them unused.
> >>>
> >>
> >> Which driver are we talking about here? The lpass clk driver? Presumably
> >> the lpass clk driver would request the GCC clks and turn them on in
> >> probe and then register any lpass clks. If the lpass clk driver probes
> >> bfeore late init, then the gcc clks will be enabled and everything
> >> works, and if the lpass clk driver probes after late init then the clks
> >> that can't be touched without gcc clks enabled won't be registered, and
> >> then they won't be touched. What goes wrong?
> >>
> >>
> > 
> > Okay, sure, I will take the GCC clock handles and then enable/disable 
> > them accordingly.
> > 
> > I missed earlier, so here is what you suggest
> 
> gcc_probe --> GCC LPASS clocks registered.
> lpass_probe --> clk_get on gcc_lpass_clocks/ clk_prepare_enable -->
> register the lpass clocks --> clk_disable_unprepare gcc_lpass_clocks.

Why did the gcc_lpass_clocks get turned off? Shouldn't they just stay
enabled all the time?

> 
> But the problem is not during the above. It is the below
> static void clk_disable_unused_subtree(struct clk_core *core)
> {
> ....
> 
>          if (clk_core_is_enabled(core)) {  --> This access fails.
> ....
> 
> }
> 

You may need to add some prepare_ops to turn on clks needed to
read/write lpass registers. Or you can look into using some sort of
genpd to enable required clks when these clks are enabled or disabled.
But I suspect it would be easier to just leave the clks in GCC for lpass
always enabled and not worry about the complicated genpd things.