[PATCH 3/4] clk: shmobile: add CPG driver for rz-platforms

[Laurent Pinchart]

Hi Wolfram,

On Monday 03 March 2014 10:19:04 Wolfram Sang wrote:
> > >  drivers/clk/shmobile/Makefile |   1 +
> > >  drivers/clk/shmobile/clk-rz.c | 112 +++++++++++++++++++++++++++++++++++
> > 
> > There's one large missing piece here: the DT bindings documentation.
> 
> Yes, noticed that, too. Added already.
> 
> > > +	u32 val;
> > > +	unsigned mult, frqcr_tab[4] = { 3, 2, 0, 1 };
> > 
> > I would declare the table as static const outside of this function.
> 
> Yup.
> 
> > > +	if (strcmp(name, "pll") == 0) {
> > > +		/* FIXME: cpg_mode should be read from GPIO. But no GPIO support
> > > yet */
> > > +		unsigned cpg_mode = 0; /* hardcoded to EXTAL for now */
> > 
> > It won't make a difference yet, but shouldn't this be moved to
> > rz_cpg_clocks_init() ?
> 
> This is going to be a gpio_get_value() later and I'd prefer it in the
> place where the value is needed.
> 
> > > +		const char *parent_name = of_clk_get_parent_name(np, 0);
> > 
> > You should select the parent depending on the mode (again it won't make
> > any difference yet, but the code will be ready, and DT bindings should
> > document two parents).
> 
> Two parents? Yet, CPG only needs one as input. Either XTAL or USB_X1
> which is board dependent. What I should do IMO is to put the parent
> property for CPG from the dtsi to the board dts. Because this is
> describing hardware. And from that parent, I can simply get its name
> like above.

While the parent is indeed selected at boot time only, and only one parent is 
thus needed, parent selection could be performed by a DIP switch connected to 
MD_CLK on the board for instance. In that case both parents should be 
available in DT, as selection will be done by the kernel at boot time, not at 
DT compile time.

> > > +	int num_clks;
> > > +
> > > +	num_clks = of_property_count_strings(np, "clock-output-names");
> > > +	if (WARN(num_clks < 0, "can't count CPG clocks\n"))
> > 
> > Do such failures really deserve a WARN ? Isn't a pr_err() enough ?
> 
> Since the system won't probably boot, I'd think so. Also, it makes the
> code more concise, no?
> 
> > What if num_clks == 0 ?
> 
> Good question.
> 
> > > +		goto out;
> > > +
> > > +	cpg = kzalloc(sizeof(*cpg), GFP_KERNEL);
> > > +	if (WARN(!cpg, "out of memory!\n"))
> > > +		goto out;
> > > +
> > > +	clks = kzalloc(num_clks * sizeof(*clks), GFP_KERNEL);
> > > +	if (WARN(!clks, "out of memory!\n"))
> > > +		goto free_cpg;
> > 
> > Does kzalloc alloc warn internally when it fails to allocate memory ?
> 
> Ehrm, why don't you simply look this up instead of asking me? :)

I have and wasn't sure :-)

> > you could remove the error message. You could also perform the two
> > allocations and check the result once only.
> 
> What is the gain? Code savings?

Yes, code saving.

> > > +free_clks:
> > > +	kfree(clks);
> > > +free_cpg:
> > > +	kfree(cpg);
> > 
> > I would remove that as done in the other CPG drivers, given that a small
> > memory leak when the system will anyway fail to boot isn't really an
> > issue.
> 
> Again, now that I already coded it, what is the gain to remove it? The
> drawback is that other people might get encouraged to find reasons to
> allow them sloppy practices.

It will make the kernel binary smaller by removing code that is not needed in 
practice.

-- 
Regards,

Laurent Pinchart
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 490 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20140303/6c45ed1e/attachment-0001.sig>