Hi Mike,

On Fri, Jul 24, 2015 at 11:26:19AM -0700, Michael Turquette wrote:

quoted

What are the issues with maintaining them? The only drawback I'm
seeing with introducing such a driver is that you can't really have a
clock that is both a divider and a multiplier, but that can be solved
by splitting it into two sub-clocks.

There are a bunch of problems with the basic clock types. First is that
there is some feature creep every merge window that subtly breaks an
existing user (e.g. the round_rate stuff in clk-divider.c), and then
there are the growing number of flags to handle corner cases that are
one-off quirks for a single chip.

These make it harder to maintain, but it is possible.

Ok. It's the downside of having common code I guess, everyone wants to
use it, and the clocks are just more subject to it than other drivers :/

The real problem with these basic clock types is that they are an
abstraction layer at the wrong level. Each clock type implements both
the policy of a given clock, as well as the machine-specific details.
For example clk-divider.c has made some assumptions in the past about
rounding the rate, or how to calculate the best divider; this is a
matter of policy and is useful on its own. But additionally that same
policy is glued to a specific implementation: memory-mapped register
controls for a clock divider.

The I/O accessor stuff needs to be addressed at some point.  Currently
the basic clock types assume specific patterns of access to
memory-mapped clock registers. There are lots of other clock controls
out there that talk to firmware, or over i2c, or whatever. The amount of
code that has to be copy/pasted for each different type of access is
100%; i.e. we do not have abstractions at the right level such as
.get_best_div(struct clk_hw *hw, unsigned long rate).

What I would like to see in time is a re-usable layer for clock policy
(e.g. common rules for how dividers or multipliers should behave), and
then have that sit on top of the machine-specific callbacks that
directly touch the hardware, such as the .get_best_div callback above.

Can't that be solved by moving to regmap using Matthias' patches, or
at least the IO method abstraction?

We would then have to only provide an additional callback then for
providers that have specific requirements about the divider
calculation.

So far, I haven't had any usecase where I needed anything but having
the as-close-as-possible rate, so I would expect to not provide
anything beside whether I'd like to round down or round up, but I do
understand that some other might have different requirements.

For this reason I like to limit the number of new basic clock types. If
there is a single user then I'm inclined to have the author put it with
the machine-specific code. But in this case since there are two users, I
see the value in making a new basic clock type.

That makes sense.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20150725/5ac50816/attachment.sig>