[RFC 1/1] clk: Add notifier support in clk_prepare_enable/clk_disable_unprepare

[Bill Huang]

On Wed, 2013-03-13 at 13:24 +0800, Stephen Warren wrote:
> On 03/12/2013 11:08 PM, Bill Huang wrote:
> > On Wed, 2013-03-13 at 12:42 +0800, Stephen Warren wrote:
> >> On 03/12/2013 07:47 PM, Bill Huang wrote:
> >>> On Tue, 2013-03-12 at 21:40 +0800, Russell King - ARM Linux wrote:
> >>>> On Tue, Mar 12, 2013 at 05:37:41AM -0700, Bill Huang wrote:
> >>>>> Add the below four notifier events so drivers which are interested in
> >>>>> knowing the clock status can act accordingly. This is extremely useful
> >>>>> in some of the DVFS (Dynamic Voltage Frequency Scaling) design.
> >>>>>
> >>>>> PRE_CLK_ENABLE
> >>>>> POST_CLK_ENABLE
> >>>>> PRE_CLK_DISABLE
> >>>>> POST_CLK_DISABLE
> >>>>>
> >>>>> Signed-off-by: Bill Huang <bilhuang at nvidia.com>
> >>>>
> >>>> NAK.  *Sigh* NO, this is the wrong level to be doing stuff like this.
> >>>>
> >>>> The *ONLY* thing that clk_prepare_enable() and clk_prepare_disable() should
> >>>> *EVER* be doing is calling clk_prepare(), clk_enable(), clk_disable() and
> >>>> clk_unprepare().  Those two functions are *merely* helpers for drivers
> >>>> who don't wish to make the individual calls.
> >>>>
> >>>> Drivers are still completely free to call the individual functions, at
> >>>> which point your proposal breaks horribly - and they _do_ call the
> >>>> individual functions.
> >>>
> >>> I'm proposing to give device driver a choice when it knows that some
> >>> driver might be interested in knowing its clock's enabled/disabled state
> >>> change at runtime, this is very important for centralized DVFS core
> >>> driver. It is not meant to be covering all cases especially for drivers
> >>> which is not part of the DVFS, so we don't care if it is calling
> >>> clk_enable/disable directly or not.
> >>
> >> I believe the point Russell is making is not that the idea behind this
> >> patch is wrong, but simply that the function where you put the hooks is
> >> wrong. The hooks should at least be in clk_enable/clk_disable and not
> >> clk_prepare_enable/clk_disable_unprepare, since any driver is free to
> >> call clk_prepare separately from clk_enable. The hooks should be
> >> implemented in the lowest-level common function that all
> >> driver-accessible paths call through.
> > 
> > Thanks, I know the point, but unfortunately there is no good choice for
> > hooking this since those low level functions clk_enable/clk_disable will
> > be called in interrupt context so it is not possible to send notify. We
> > might need to come out a better approach if we can think of any.
> > Currently I still think this is acceptable (Having all the drivers which
> > are using our interested clocks call these function to enable/disable
> > clock in their runtime_pm calls) though it's not perfect. 
> 
> No, that definitely won't work. Not all drivers use those APIs, nor
> should they.
> 
That will be too bad, it looks like we deadlock in the mechanism, we
cannot change existing drivers behavior (that means some call
clk_disable/enable directly, some are not), and we cannot hook notifier
in clk_disable/enable either, that means there seems no any chance to
get what we want, any idea?