Re: [PATCH v8 03/10] clk: eyeq5: add platform driver, and init routine at of_clk_init()

[Théo Lebrun]

Hello,

On Thu Feb 29, 2024 at 3:59 PM CET, Andy Shevchenko wrote:
> On Thu, Feb 29, 2024 at 03:27:01PM +0100, Théo Lebrun wrote:
> > On Wed, Feb 28, 2024 at 03:33:29PM +0100, Théo Lebrun wrote:
> > > On Tue Feb 27, 2024 at 6:11 PM CET, Andy Shevchenko wrote:
> > > > On Tue, Feb 27, 2024 at 03:55:24PM +0100, Théo Lebrun wrote:
>
> [...]
>
> > > > > > +	u32		reg;	/* next 8 bytes are r0 and r1 */
> > > > >
> > > > > Not sure this comments gives any clarification to a mere reader of the code.
> > > > > Perhaps you want to name this as reg64 (at least it will show that you have
> > > > > 8 bytes, but I have no clue what is the semantic relationship between r0 and
> > > > > r1, it's quite cryptic to me). Or maybe it should be reg_0_1?
> > > > 
> > > > Clocks are defined by two 32-bit registers. We only store the first
> > > > register offset because they always follow each other.
> > >
> > > > I like the reg64 name and will remove the comment. This straight forward
> > > > code is found in the rest of the code, I don't think it is anything
> > > > hard to understand (ie does not need a comment):
> > > > 
> > > > 	u32 r0 = readl(base_plls + pll->reg);
> > > > 	u32 r1 = readl(base_plls + pll->reg + sizeof(r0));
> > >
> > > Btw, why readq()/writeq() (with probably the inclusion of io-64-nonatomic-lo-hi.h)
> > > can be used in this case? It will be much better overall and be aligned with
> > > reg64 name.
> > 
> > The doc talks in terms of 32-bit registers. I do not see a reason to
> > work in 64-bit. If we get a 64-bit value that we need to split we need
> > to think about the endianness of our platform, which makes things more
> > complex than just reading both values independently.
>
> 1) Would be nice to test on the real HW to confirm it doesn't accept 64-bit IO.

Just tested, it works. No error on the memory bus. And checked assembly
generated was a single 64-bit instructions.

It might not work on other hardware revisions though. I can't remember
if memory bus is changing across them.

> 2) Still I see a benefit from using lo_hi_readq() and friends directly.

So it is:

	u32 r0 = readl(base_plls + pll->reg64);
	u32 r1 = readl(base_plls + pll->reg64 + sizeof(r0));

vs:

	u64 r = lo_hi_readq(base_plls + pll->regs64);
	u32 r0 = r;
	u32 r1 = r >> 32;

One is straight forward, the other uses an obscure helper that code
readers must understand and follows that with bit manipulation.

>
> [...]
>
> > > > > I didn't get. If eq5c_init() was finished successfully, why do you need to
> > > > > seems repeat what it already done? What did I miss?
> > > > 
> > > > The key here is that eq5c_init() iterates on eq5c_early_plls[] while
> > > > eq5c_probe() iterates on eq5c_plls[]. I've tried to hint at this in the
> > > > commit message:
> > > > 
> > > > > Two PLLs are required early on and are therefore registered at
> > > > > of_clk_init(). Those are pll-cpu for the GIC timer and pll-per for the
> > > > > UARTs.
> > > > 
> > > > Doing everything in eq5c_init() is not clean because we expect all new
> > > > clock provider drivers to be standard platform drivers. Doing
> > > > everything from a platform driver probe doesn't work because some
> > > > clocks are required earlier than platform bus init. We therefore do a
> > > > mix.
> > >
> > > Am I missing something or these two pieces are using the same IO resources?
> > > This looks like a lot of code duplication without clear benefit. Perhaps
> > > you can have a helper?
> > 
> > There are two subtle differences that make creating a helper difficult:
> > 
> >  - Logging, pr_*() vs dev_*(). Second option is preferred but only
> >    available once a device is created.
>
> Some code uses (yeah, arguable that it's better, but depends on how much
> the real deduplication takes)
>
> 	if (dev)
> 		dev_*(...);
> 	else
> 		pr_*(...);
>
> >  - Behavior on error: we stop the world for early clocks but keep going
> >    for normal clocks.
>
> ...(..., bool skip_errors)
> {
> 	...
> }
>
> (with the same caveat)?

I started trying it out, but the combination of both flags means dealing
with errors would look like:

	ret = foo();
	if (ret) {
		if (!skip_errors) {
			if (dev)
				dev_err(dev, "...");
			else
				pr_err("...");
			return ret;
		}
		if (dev)
			dev_warn(dev, "...");
		else
			pr_warn("...");
	}

There are two errors to handle, that makes a mess out of the code.
Having a little bit of repetition but straight forward code is nicer in
my opinion. At least we tried!

Regards,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com