Re: [PATCH v2 1/2] clk: change clk_ops' ->round_rate() prototype

[Boris Brezillon <boris.brezillon@free-electrons.com>]

Hi Stephen,

On Wed, 6 May 2015 23:39:53 -0700
Stephen Boyd <sboyd@codeaurora.org> wrote:

> On 04/30, Boris Brezillon wrote:
> > Clock rates are stored in an unsigned long field, but ->round_rate()
> > (which returns a rounded rate from a requested one) returns a long
> > value (errors are reported using negative error codes), which can lead
> > to long overflow if the clock rate exceed 2Ghz.
> > 
> > Change ->round_rate() prototype to return 0 or an error code, and pass the
> > requested rate as a pointer so that it can be adjusted depending on
> > hardware capabilities.
> > 
> > Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>
> > Tested-by: Heiko Stuebner <heiko@sntech.de>
> > Tested-by: Mikko Perttunen <mikko.perttunen@kapsi.fi>
> > Reviewed-by: Heiko Stuebner <heiko@sntech.de>
> 
> This patch is fairly invasive, and it probably doesn't even
> matter for most of these clock providers to be able to round a
> rate above 2GHz.

Fair enough.

> I've been trying to remove the .round_rate op
> from the framework by encouraging new features via the
> .determine_rate op.

Oh, I wasn't aware of that (BTW, that's a good thing).
Maybe this should be clearly stated (both in the struct clk_ops
kerneldoc header and in Documentation/clk.txt).

> Sadly, we still have to do a flag day and
> change all the .determine_rate ops when we want to add things.

Yes, but the number of clk drivers implementing ->determine_rate() is
still quite limited compared to those implementing ->round_rate().

> 
> What if we changed determine_rate ops to take a struct
> clk_determine_info (or some better named structure) instead of
> the current list of arguments that it currently takes? Then when
> we want to make these sorts of framework wide changes we can just
> throw a new member into that structure and be done.

I really like this idea, especially since I was wondering if we could
pass other 'clk rate requirements' like the rounding policy (down,
closest, up), or the maximum clk inaccuracy.

> 
> It doesn't solve the unsigned long to int return value problem
> though. We can solve that by gradually introducing a new op and
> handling another case in the rounding path. If we can come up
> with some good name for that new op like .decide_rate or
> something then it makes things nicer in the long run. I like the
> name .determine_rate though :/

Why not changing the ->determine_rate() prototype. As said above, the
number of clk drivers implementing this function is still quite
limited, and I guess we can have an ack for all of them.

> 
> The benefit of all this is that we don't have to worry about
> finding the random clk providers that get added into other
> subsystems and fixing them up. If drivers actually care about
> this problem then they'll be fixed to use the proper op. FYI,
> last time we updated the function signature of .determine_rate we
> broke a couple drivers along the way.
> 

Hm, IMHO, adding a new op is not a good thing. I agree that it eases
the transition, but ITOH you'll have to live with old/deprecated ops in
your clk_ops structure with people introducing new drivers still using
the old ops (see the number of clk drivers implementing ->round_rate()
instead of ->determine_rate()).

Best Regards,

Boris

-- 
Boris Brezillon, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com