On Fri, Mar 23, 2012 at 3:12 PM, Saravana Kannan [off-list ref] wrote:

On 03/23/2012 02:39 PM, Turquette, Mike wrote:

quoted

__clk_recalc_rates is called by __clk_reparent which is called by
clk_set_parent. ?__clk_recalc_rates is also called by clk_set_rate.

Does this not handle the old cached clk->rate for you?

For the set_parent case, ops->recalc_rate() is called twice. Once for
PRE_CHANGE and once for POST_CHANGE. For this clock, I can only really
recalc the rate during the POST_CHANGE call. So, how should I differentiate
the two cases?

.recalc_rate serves two purposes: first it recalculates the rate after
the rate has changed and you pass in a new parent_rate argument.  The
second purpose is to "speculate" a rate change.  You can pass in any
rate for the parent_rate parameter when you call .recalc_rates.  This
is what __speculate_rates does before the rate changes.  For
clk_set_parent we call,

__clk_speculate_rates(clk, parent->rate);

Where parent above is the *new* parent.  So this will let us propagate
pre-rate change notifiers with the new rate.

Your .recalc_rate callback doesn't need to differentiate between the
two calls to .recalc_rate.  It should just take in the parent_rate
value and do the calculation required of it.

Take a look at __clk_speculate_rates and __clk_recalc_rates and you'll
see how to use it.

On a separate note:
Sorry if I missed any earlier discussion on this, but what's the reason for
calling recalc_rate() pre-change and post-change but without giving it the
ability to differentiate between the two?

I think my answer above covers this.  The .recalc_rate callback only
exists to perform a hardware-specific calculation.  The context of
pre- or post-change is irrelevant since the parent_rate passed in
could be the actual rate of the parent or a future rate of the parent.

I think it's quite useful for recalc_rate to be called pre/post change (some
steps have to be done pre/post change depending on whether the parent rate
is increasing or decreasing). But I don't see the "msg" being passed along.

What kind of steps?  Does your .recalc_rate perform these steps?  I
need more details to understand your requirements.

Also, I noticed that clk_set_parent() is treating a NULL as an invalid
clock. Should that be fixed? set_parent(NULL) could be treated as a
grounding the clock. Should we let the ops->set_parent determine if NULL is
valid option?

We must be looking at different code.  clk_set_parent doesn't return
any error if parent == NULL.  Bringing this to my attention does show
that we do deref the parent pointer without a check though...

Do you have a real use case for this?  Due to the way that we match
the parent pointer with the cached clk->parents member it would be
painful to support NULL parents as valid.

It is also worth considering whether clk_set_parent is really the
correct operation for grounding a clock.  clk_unprepare might be a
better candidate.

Regards,
Mike