Re: [PATCH] clk: qcom: gcc-sdm660: Replace usage of parent_names

[Bjorn Andersson]

On Wed 25 Aug 11:00 CDT 2021, Marijn Suijten wrote:

> Hi Bjorn,
> 
> On 8/25/21 12:38 AM, Bjorn Andersson wrote:
> > On Tue 24 Aug 13:46 PDT 2021, Marijn Suijten wrote:
> > 
> > > Hi Bjorn,
> > > 
> > > Thanks for this cleanup, that's needed and much appreciated!
> > > 
> > > On 8/24/21 5:06 PM, Bjorn Andersson wrote:
> > > > Using parent_data and parent_hws, instead of parent_names, does protect
> > > > against some cases of incompletely defined clock trees. While it turns
> > > > out that the bug being chased this time was totally unrelated, this
> > > > patch converts the SDM660 GCC driver to avoid such issues.
> > > > 
> > > > Signed-off-by: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx>
> > > 
> > > 
> > > Tested-by: Marijn Suijten <marijn.suijten@xxxxxxxxxxxxxx>
> > > 
> > > On the Sony Xperia XA2 Ultra, bar the necessary change in the 14NM DSI PHY
> > > driver commented below.
> > > 
> > > > [..]
> > > > -
> > > > -static struct clk_fixed_factor xo = {
> > > > -	.mult = 1,
> > > > -	.div = 1,
> > > > -	.hw.init = &(struct clk_init_data){
> > > > -		.name = "xo",
> > > > -		.parent_names = (const char *[]){ "xo_board" },
> > > > -		.num_parents = 1,
> > > > -		.ops = &clk_fixed_factor_ops,
> > > > -	},
> > > > -};
> > > 
> > > 
> > > Removing the global "xo" clock makes it so that our 14nm DSI PHY does not
> > > have a parent clock anymore, as the clock is called "xo_board" nowadays
> > > ("xo" in the position of fw_name is, as you know, only local to this driver
> > > because it is named that way in the clock-names property). We (SoMainline)
> > > suffer the same DSI PHY hardcoding issue on many other boards and are at
> > > this point investigating whether to provide &xo_board in DT like any other
> > > sane driver.  Do you happen to know if work is already underway to tackle
> > > this?
> > > 
> > 
> > As far as I can tell most other platforms doesn't define "xo" either.
> > E.g. according to debugfs dsi0vco_clk doesn't have a parent on sdm845...
> > 
> > Sounds like we should update the dsi phys to specify a fw_name and
> > update binding and dts to provide this...
> 
> 
> I'm all for using .fw_name there, and I hope we all agree that clock
> dependencies based on global names should become a thing of the past; every
> such inter-driver dependency should be clearly visible in the DT.  We
> (SoMainline) can tackle this DSI side if no-one else is working on it yet.
> 
> > Does this cause a noticeable regression or it's just that we have a
> > dangling clock?
> 
> 
> Unfortunately this regresses yes, starting with:
> 
>     dsi0n1_postdiv_clk: Zero divisor and CLK_DIVIDER_ALLOW_ZERO not set
> 
> And proceeding with more such errors on different clocks, clocks getting
> stuck or failing to update, and the panel never showing anything at all.
> 
> Should we fix DSI PHYs first and let this patch sit for a while, or keep the
> implicit global "xo" clock just a little while longer until that's over
> with?
> 

Thanks, should have read your email as well before replying to the
other.

I will respin this with "xo" intact, then once we've fixed up the DSI
code we can drop it.


But I still don't understand why we don't have this problem on e.g.
sdm845. I must be missing something...

> Either way, feel free to attach my:
> 
> Reviewed-by: Marijn Suijten <marijn.suijten@xxxxxxxxxxxxxx>
> 
> After that.
> 
> > > >    static struct clk_alpha_pll gpll0_early = {
> > > >    	.offset = 0x0,
> > > >    	.regs = clk_alpha_pll_regs[CLK_ALPHA_PLL_TYPE_DEFAULT],
> > > > @@ -158,7 +35,9 @@ static struct clk_alpha_pll gpll0_early = {
> > > >    		.enable_mask = BIT(0),
> > > >    		.hw.init = &(struct clk_init_data){
> > > >    			.name = "gpll0_early",
> > > > -			.parent_names = (const char *[]){ "xo" },
> > > > +			.parent_data = &(const struct clk_parent_data){
> > > > +				.fw_name = "xo",
> > > > +			},
> > > 
> > > 
> > > I wish we could use .parent_names for a list of .fw_name's too
> > 
> > Afaict specifying "name" in struct clk_parent_data is the same as using
> > parent_names. But I'm not up to speed on the details of how to migrate
> > the dsi phys.
> 
> 
> Yes it is, both do _not_ look at clocks specified in DT before "falling
> back" to global names (that only happens when both .name and .fw_name are
> specified).  I'm sort of expressing the desire for .parent_fw_names here in
> hopes of phasing out global clock names on DT platforms altogether.  We
> definitely shouldn't rework .parent_names to support both, that only causes
> confusion and an implicit fallback to global clocks when the DT is
> under-specifying the required clocks is exactly what we're trying to avoid.
> 
> > > > [..]
> > > > @@ -265,7 +270,7 @@ static struct clk_rcg2 blsp1_qup1_i2c_apps_clk_src = {
> > > >    	.freq_tbl = ftbl_blsp1_qup1_i2c_apps_clk_src,
> > > >    	.clkr.hw.init = &(struct clk_init_data){
> > > >    		.name = "blsp1_qup1_i2c_apps_clk_src",
> > > > -		.parent_names = gcc_parent_names_xo_gpll0_gpll0_early_div,
> > > > +		.parent_data = gcc_parent_data_xo_gpll0_gpll0_early_div,
> > > >    		.num_parents = 3,
> > > 
> > > 
> > > How about using ARRAY_SIZE(gcc_parent_data_xo_gpll0_gpll0_early_div) now?
> > > Same for every other occurrence of this pattern.
> > > 
> > 
> > I omitted that because it felt unrelated to the change I was doing, but
> > it could certainly be done.
> 
> 
> Fair, if done at all it should end up in a separate (2/2) patch or I'll take
> care of this in a followup.
> 

Sounds good, I'd be happy to give you a review on that patch :)

Regards,
Bjorn