Re: [PATCH RFC RFT 0/3] clk: detect per-user enable imbalances and implement hand-off

[Lee Jones]

On Mon, 10 Aug 2015, Michael Turquette wrote:
> Quoting Lee Jones (2015-08-10 08:36:38)
> > On Fri, 07 Aug 2015, Michael Turquette wrote:
> > > This is an alternative solution to Lee's "clk: Provide support for
> > > always-on clocks" series[0].
> >  
> > So after all the hours of work that's been put in and all of the
> > versions that have been authored, you're just going to write your own
> > solution anyway, without any further discussion?
> 
> This certainly seems like "further discussion" to me. In fact the
> capital C in RFC pretty much announces that this is a big, wide open
> discussion.
> 
> If I had ninja-merged the patches then I would understand your
> frustration, but that clearly is not the case.

This doesn't have anything to do with merging code.

When contributors submit patches to subsystem I maintain, I provide
reviews, advice and guidance.  If I think an approach is wrong, I
state why I think that and provide ideas for an alternative approach
if required.  If instead of providing feedback I just replied with a
patch-set containing my own alternative method, I would expect the
original submitter to be justifiably cross.

> > That's one way to make your contributors feel valued.
> 
> Lee, your contributions are very much valued. You and I both know that
> there are times when a concept is far better communicated with code than
> with prose. This is one of those times. The DT-centric solution that
> requires a clock consumer driver to call clk_disable() before calling
> clk_enable() (an obvious violation of the clk.h api) is just plain wrong
> and this implementation hopes to get it back on the right course.

See my previous email about what I think about "bastardizing the clk.h
api".

> > > The first two patches introduce run-time checks to ensure that clock
> > > consumer drivers are respecting the clk.h api. The former patch checks
> > > for prepare and enable imbalances. The latter checks for calls to
> > > clk_put without first disabling and unpreparing the clk.
> > 
> > Are these real issues that people are having trouble with, or just
> > hypothetical ones?  I can understand the need for them if they're
> > causing people some pain, but if they are unnecessarily hardening the
> > API, then it makes it difficult for proper issues (such as critical
> > clocks) to be solved easily.
> 
> They are deeply related. Per-user accounting gives us a graceful method
> to implement hand-off, which is the real feature that is needed here.

The hand-off feature was mentioned a week ago, as part of an 8 month
discussion.  This only aids knowledgeable consumers to turn of a
critical clock after the introduction of the API tightening which
happens in _this_ set.  So I disagree that this is the "real feature
that is needed here".

The _real_ feature that's required is a generic way to keep critical
clocks enabled and a way for them to be gated if another driver knows
best.  Granted this method is pretty clean for drivers which have all
of their platform data in driver code, but it's still not ideal.

> > > The third patch introduces a new flag, CLK_ENABLE_HAND_OFF, which
> > > prepares and enables a clk at registration-time. The reference counts
> > > (prepare & enable) are transferred to the first clock consumer driver
> > > that clk_get's the clk with this flag set AND calls clk_prepare or
> > > clk_enable.
> > > 
> > > The net result is that a clock with this flag set will be enabled at
> > > boot and neither the clk_disable_unused garbage collector or the
> > > "sibling clock disables a shared parent" scenario will cause the flagged
> > > clock to be disabled.
> > 
> > I'm not sure if the third patch actually solves the _real_ issue you
> > allude to here.  The original critical (then called always-on) clock
> > patch-set solved it just fine.  However others were concerned about
> > how you would then turn the clock off if some knowledgeable consumer
> > (who knew the full consequences of their actions) came along and had a
> > good reason to gate it.  That, the turning off the clock if still
> > desired is what the third patch _actually_ solves.
> > 
> > Now we are back where we started however, and still need to figure out
> > a generic method to mark the clocks (either by setting a FLAG or
> > actually calling enable()) as critical.
> 
> The third patch:
> 
> 1) implements an always-on behavior
> 2) allows knowledgeable drivers to gate the clock
> 3) marks such clocks with a flag
> 
> What is missing?

The difficuly is marking the clock in the first place.  We already
have drivers doing 1), just not in a generic way.  This solution is
more generic (as was mine), but it's still not completely generic.

2) is a new requirement that would have required a little more
thought.  My solution to this was only in V1/RFC and admitted would
have required a little more discussion.

3) isn't really a 'thing'.

> > > The first driver to come along and explicitly
> > > claim, prepare and enable this clock will inherit those reference
> > > counts. No change to clock consumer drivers is required for this to
> > > work.
> > 
> > Do you mean it won't break anything?  I think to make use of the
> > functionality each of the providers still have to figure out which of
> > their clocks are critical (which may change from platform to platform)
> > then mark them as such before registration.
> 
> What I meant was: "does not require any new api such as
> clk_critical_enable() or clk_critical_disable()". There is zero change
> to clk.h as a part of this series, which is a good thing.

Okay, fine.

> > > Please continue to use the clk.h api properly.
> > > 
> > > In time this approach can probably replace the CLK_IGNORE_UNUSED flag
> > > and hopefully reduce the number of users of the clk_ignore_unused boot
> > > parameter.
> > > 
> > > Finally, a quick note on comparing this series to Lee's. I went with the
> > > simplest approach to solve a real problem: preventing critical clocks
> > > from being spuriously disabled at boot, or before a their parent clock
> > > becomes accidentally disabled by a sibling.
> > 
> > This wasn't the problem.  Platforms are already doing this.  The _real_
> 
> Yes, this is a very real problem. In fact it is two very real problems:
> 
> 1) Platforms are open-coding this "always-on" behavior by calling
> clk_prepare_enable a bunch in their clock provider driver. I'm still OK
> with that, but this flag provides a coherent way to do it. And,
> 
> 2) Drivers that open-code calls to clk_prepare_enable in their clock
> provider driver have no way to allow clock knowledgeable consumer
> drivers to gate those clocks later on.
> 
> > problem was doing it in a generic way, so vendors didn't have to roll
> > their own 'gather' and 'mark' code, which unfortunately they still have
> > to do with this solution.
> 
> There is no gather. The data for clocks belongs in most drivers, just
> not yours. Marking is as simple as setting a flag (which many drives
> already do). Please see my response to you in patch #3 for an example.

I have also replied in #3.  No need for me to labour the points here
too.

> > > All of the other kitchen sink stuff (DT binding, 
> > 
> > The DT binding will still be required, at least for ST, as they have
> > gone for the more Linuxy approach of having a generic set of clock
> > drivers and provide all of the platform specific information in
> > platform code (i.e. DT).  So to identify which clocks are critical on
> > each platform the DT will need to be interrogated in some way.
> 
> At the risk of instigating a serious religious conflict, I humbly
> disagree that using Devicetree as a data-driven interface for device
> drivers is the "more Linuxy approach". I think that the Devicetree
> people would disagree with you too.
> 
> The more Linuxy approach is for device drivers to contain all of the
> information they need to function.  Devicetree does a stellar job of
> linking up providers and consumers of these resources, which is outside
> the purview of individual drivers.

I agree to disagree.

> > > passing the flag back to the framework when the clock consumer
> > > driver calls clk_put) was left
> > 
> > I'm not sure what this is.
> > 
> > > out because I do not see a real use case for it. If one can demonstrate
> > > a real use case (and not a hypothetical one) then this patch series can
> > > be expanded further.
> > > 
> > > [0] http://lkml.kernel.org/r/<1437570255-21049-1-git-send-email-lee.jones@xxxxxxxxxx>
> > > 
> > > Michael Turquette (3):
> > >   clk: per-user clk prepare & enable ref counts
> > >   clk: clk_put WARNs if user has not disabled clk
> > >   clk: introduce CLK_ENABLE_HAND_OFF flag
> > > 
> > >  drivers/clk/clk.c            | 79 +++++++++++++++++++++++++++++++++++++++++---
> > >  include/linux/clk-provider.h |  3 ++
> > >  2 files changed, 78 insertions(+), 4 deletions(-)
> > > 
> > 

-- 
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org â Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/