Hi,
On Thu, Jun 18, 2020 at 9:55 AM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
>
> Quoting Doug Anderson (2020-06-18 08:32:20)
> > Hi,
> >
> > On Thu, Jun 18, 2020 at 7:01 AM Srinivas Kandagatla
> > >
> > > On the other note:
> > >
> > > clock-names are not mandatory according to
> > > Documentation/devicetree/bindings/clock/clock-bindings.txt
> > >
> > > For this particular case where clock-names = "sec" is totally used for
> > > indexing and nothing else!
> >
> > So I guess in the one-clock case it's more optional and if you feel
> > strongly I'll get rid of clk-names here. ...but if we ever need
> > another clock we probably will want to add it back and (I could be
> > corrected) I believe it's convention to specify clk-names even with
> > one clock.
>
> TL;DR: I suggest you call this "core" if you want to keep the
> clock-name, or just drop it if there's only one clk and move on.
Ah, true. "core" sounds good.
> It's not required to have clock-names with one clk, and indeed it's not
> required to have clock-names at all. The multi clk scenario is a little
> more difficult to handle because historically the clk_get() API has been
> name based and not index based like platform resources. When there is
> one clk the driver can pass NULL as the 'con_id' argument to clk_get()
> and it will do the right thing. And when you have more than one clk you
> can pass NULL still and get the first clk, that should be in the same
> index, and then other clks by name.
>
> So far nobody has added clk_get_by_index() but I suppose if it was
> important the API could be added. Working with only legacy clkdev
> lookups would fail of course, but clock-names could be fully deprecated
> and kernel images may be smaller because we're not storing piles of
> strings and doing string comparisons. Given that it's been this way for
> a long time and we have DT schema checking it doesn't seem very
> important to mandate anything one way or the other though. I certainly
> don't feel good when I see of_clk_*() APIs being used by platform
> drivers, but sometimes it is required.
>
> To really put this into perspective, consider the fact that most drivers
> have code that figures out what clk names to look for and then they pile
> them into arrays and just turn them all on and off together. Providing
> fine grained clk control here is a gigantic waste of time, and requiring
> clock-names is just more hoops that driver authors feel they have to
> jump through for $reasons. We have clk_bulk_get_all() for this, but that
> doesn't solve the one rate changing clk among the sea of clk gates
> problem. In general, driver authors don't care and we should probably be
> providing a richer while simpler API to them that manages power state of
> some handful of clks, regulators, and power domains for a device while
> also letting them control various knobs like clk rate when necessary.
>
> BTW, on qcom platforms they usually name clks "core" and "iface" for the
> core clk and the interface clk used to access the registers of a device.
> Sometimes there are esoteric ones like "axi". In theory this cuts down
> on the number of strings the kernel keeps around but I like that it
> helps provide continuity across drivers and DTs for their SoCs. If you
> ask the hardware engineer what the clk name is for the hardware block
> they'll tell you the globally unique clk name like
> "gcc_qupv3_uart9_core_clk", which is the worst name to use.
OK, sounds about what I expected. I suppose the path of least
resistance would be to just drop clock-names. I guess I'm just
worried that down the road someone will want to specify the "iface"
clock too. If that ever happens, we're stuck with these options:
1. Be the first ones to require adding clk_get_by_index().
2. Use the frowned upon of_clk_get() API which allows getting by index.
3. Get the first clock with clk_get(NULL) and the second clock with
clk_get("iface") and figure out how to specify this happily in the
yaml.
If we just define clock-names now then we pretty much match the
pattern of everyone else.
Srinivas: reading all that if you still want me to drop clock-names
then I will. I'll use clk_get(NULL) to get the clock and if/when we
ever need an "iface" clock (maybe we never will?) we can figure it out
then.
-Doug
