Re: [PATCH v5 05/12] PM / devfreq: Add support for policy notifiers

[Matthias Kaehlcke]

Hi Chanwoo,

On Tue, Aug 07, 2018 at 10:35:37AM +0900, Chanwoo Choi wrote:
> Hi Matthias,
> 
> On 2018ë 08ì 07ì 09:23, Matthias Kaehlcke wrote:
> > Hi Chanwoo,
> > 
> > On Tue, Aug 07, 2018 at 07:31:16AM +0900, Chanwoo Choi wrote:
> >> Hi Matthias,
> >>
> >> On 2018ë 08ì 07ì 04:21, Matthias Kaehlcke wrote:
> >>> Hi Chanwoo,
> >>>
> >>> On Fri, Aug 03, 2018 at 09:14:46AM +0900, Chanwoo Choi wrote:
> >>>> Hi Matthias,
> >>>>
> >>>> On 2018ë 08ì 03ì 08:48, Matthias Kaehlcke wrote:
> >>>>> On Thu, Aug 02, 2018 at 04:13:43PM -0700, Matthias Kaehlcke wrote:
> >>>>>> Hi Chanwoo,
> >>>>>>
> >>>>>> On Thu, Aug 02, 2018 at 10:58:59AM +0900, Chanwoo Choi wrote:
> >>>>>>> Hi Matthias,
> >>>>>>>
> >>>>>>> On 2018ë 08ì 02ì 02:08, Matthias Kaehlcke wrote:
> >>>>>>>> Hi Chanwoo,
> >>>>>>>>
> >>>>>>>> On Wed, Aug 01, 2018 at 10:22:16AM +0900, Chanwoo Choi wrote:
> >>>>>>>>> On 2018ë 08ì 01ì 04:39, Matthias Kaehlcke wrote:
> >>>>>>>>>> On Mon, Jul 16, 2018 at 10:50:50AM -0700, Matthias Kaehlcke wrote:
> >>>>>>>>>>> On Thu, Jul 12, 2018 at 05:44:33PM +0900, Chanwoo Choi wrote:
> >>>>>>>>>>>> Hi Matthias,
> >>>>>>>>>>>>
> >>>>>>>>>>>> On 2018ë 07ì 07ì 02:53, Matthias Kaehlcke wrote:
> >>>>>>>>>>>>> Hi Chanwoo,
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> On Wed, Jul 04, 2018 at 03:41:46PM +0900, Chanwoo Choi wrote:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>> Firstly,
> >>>>>>>>>>>>>> I'm not sure why devfreq needs the devfreq_verify_within_limits() function.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> devfreq already used the OPP interface as default. It means that
> >>>>>>>>>>>>>> the outside of 'drivers/devfreq' can disable/enable the frequency
> >>>>>>>>>>>>>> such as drivers/thermal/devfreq_cooling.c. Also, when some device
> >>>>>>>>>>>>>> drivers disable/enable the specific frequency, the devfreq core
> >>>>>>>>>>>>>> consider them.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> So, devfreq doesn't need to devfreq_verify_within_limits() because
> >>>>>>>>>>>>>> already support some interface to change the minimum/maximum frequency
> >>>>>>>>>>>>>> of devfreq device. 
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> In case of cpufreq subsystem, cpufreq only provides 'cpufreq_verify_with_limits()'
> >>>>>>>>>>>>>> to change the minimum/maximum frequency of cpu. some device driver cannot
> >>>>>>>>>>>>>> change the minimum/maximum frequency through OPP interface.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> But, in case of devfreq subsystem, as I explained already, devfreq support
> >>>>>>>>>>>>>> the OPP interface as default way. devfreq subsystem doesn't need to add
> >>>>>>>>>>>>>> other way to change the minimum/maximum frequency.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> Using the OPP interface exclusively works as long as a
> >>>>>>>>>>>>> enabling/disabling of OPPs is limited to a single driver
> >>>>>>>>>>>>> (drivers/thermal/devfreq_cooling.c). When multiple drivers are
> >>>>>>>>>>>>> involved you need a way to resolve conflicts, that's the purpose of
> >>>>>>>>>>>>> devfreq_verify_within_limits(). Please let me know if there are
> >>>>>>>>>>>>> existing mechanisms for conflict resolution that I overlooked.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> Possibly drivers/thermal/devfreq_cooling.c could be migrated to use
> >>>>>>>>>>>>> devfreq_verify_within_limits() instead of the OPP interface if
> >>>>>>>>>>>>> desired, however this seems beyond the scope of this series.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Actually, if we uses this approach, it doesn't support the multiple drivers too.
> >>>>>>>>>>>> If non throttler drivers uses devfreq_verify_within_limits(), the conflict
> >>>>>>>>>>>> happen.
> >>>>>>>>>>>
> >>>>>>>>>>> As long as drivers limit the max freq there is no conflict, the lowest
> >>>>>>>>>>> max freq wins. I expect this to be the usual case, apparently it
> >>>>>>>>>>> worked for cpufreq for 10+ years.
> >>>>>>>>>>>
> >>>>>>>>>>> However it is correct that there would be a conflict if a driver
> >>>>>>>>>>> requests a min freq that is higher than the max freq requested by
> >>>>>>>>>>> another. In this case devfreq_verify_within_limits() resolves the
> >>>>>>>>>>> conflict by raising p->max to the min freq. Not sure if this is
> >>>>>>>>>>> something that would ever occur in practice though.
> >>>>>>>>>>>
> >>>>>>>>>>> If we are really concerned about this case it would also be an option
> >>>>>>>>>>> to limit the adjustment to the max frequency.
> >>>>>>>>>>>
> >>>>>>>>>>>> To resolve the conflict for multiple device driver, maybe OPP interface
> >>>>>>>>>>>> have to support 'usage_count' such as clk_enable/disable().
> >>>>>>>>>>>
> >>>>>>>>>>> This would require supporting negative usage count values, since a OPP
> >>>>>>>>>>> should not be enabled if e.g. thermal enables it but the throttler
> >>>>>>>>>>> disabled it or viceversa.
> >>>>>>>>>>>
> >>>>>>>>>>> Theoretically there could also be conflicts, like one driver disabling
> >>>>>>>>>>> the higher OPPs and another the lower ones, with the outcome of all
> >>>>>>>>>>> OPPs being disabled, which would be a more drastic conflict resolution
> >>>>>>>>>>> than that of devfreq_verify_within_limits().
> >>>>>>>>>>>
> >>>>>>>>>>> Viresh, what do you think about an OPP usage count?
> >>>>>>>>>>
> >>>>>>>>>> Ping, can we try to reach a conclusion on this or at least keep the
> >>>>>>>>>> discussion going?
> >>>>>>>>>>
> >>>>>>>>>> Not that it matters, but my preferred solution continues to be
> >>>>>>>>>> devfreq_verify_within_limits(). It solves conflicts in some way (which
> >>>>>>>>>> could be adjusted if needed) and has proven to work in practice for
> >>>>>>>>>> 10+ years in a very similar sub-system.
> >>>>>>>>>
> >>>>>>>>> It is not true. Current cpufreq subsystem doesn't support external OPP
> >>>>>>>>> control to enable/disable the OPP entry. If some device driver
> >>>>>>>>> controls the OPP entry of cpufreq driver with opp_disable/enable(),
> >>>>>>>>> the operation is not working. Because cpufreq considers the limit
> >>>>>>>>> through 'cpufreq_verify_with_limits()' only.
> >>>>>>>>
> >>>>>>>> Ok, we can probably agree that using cpufreq_verify_with_limits()
> >>>>>>>> exclusively seems to have worked well for cpufreq, and that in their
> >>>>>>>> overall purpose cpufreq and devfreq are similar subsystems.
> >>>>>>>>
> >>>>>>>> The current throttler series with devfreq_verify_within_limits() takes
> >>>>>>>> the enabled OPPs into account, the lowest and highest OPP are used as
> >>>>>>>> a starting point for the frequency adjustment and (in theory) the
> >>>>>>>> frequency range should only be narrowed by
> >>>>>>>> devfreq_verify_within_limits().
> >>>>>>>>
> >>>>>>>>> As I already commented[1], there is different between cpufreq and devfreq.
> >>>>>>>>> [1] https://lkml.org/lkml/2018/7/4/80
> >>>>>>>>>
> >>>>>>>>> Already, subsystem already used OPP interface in order to control
> >>>>>>>>> specific OPP entry. I don't want to provide two outside method
> >>>>>>>>> to control the frequency of devfreq driver. It might make the confusion.
> >>>>>>>>
> >>>>>>>> I understand your point, it would indeed be preferable to have a
> >>>>>>>> single method. However I'm not convinced that the OPP interface is
> >>>>>>>> a suitable solution, as I exposed earlier in this thread (quoted
> >>>>>>>> below).
> >>>>>>>>
> >>>>>>>> I would like you to at least consider the possibility of changing
> >>>>>>>> drivers/thermal/devfreq_cooling.c to devfreq_verify_within_limits().
> >>>>>>>> Besides that it's not what is currently used, do you see any technical
> >>>>>>>> concerns that would make devfreq_verify_within_limits() an unsuitable
> >>>>>>>> or inferior solution?
> >>>>>>>
> >>>>>>> As we already discussed, devfreq_verify_within_limits() doesn't support
> >>>>>>> the multiple outside controllers (e.g., devfreq-cooling.c).
> >>>>>>
> >>>>>> That's incorrect, its purpose is precisely that.
> >>>>>>
> >>>>>> Are you suggesting that cpufreq with its use of
> >>>>>> cpufreq_verify_within_limits() (the inspiration for
> >>>>>> devfreq_verify_within_limits()) is broken? It is used by cpu_cooling.c
> >>>>>> and other drivers when receiving a CPUFREQ_ADJUST event, essentially
> >>>>>> what I am proposing with DEVFREQ_ADJUST.
> >>>>>>
> >>>>>> Could you elaborate why this model wouldn't work for devfreq? "OPP
> >>>>>> interface is mandatory for devfreq" isn't really a technical argument,
> >>>>>> is it mandatory for any other reason than that it is the interface
> >>>>>> that is currently used?
> >>>>>>
> >>>>>>> After you are suggesting the throttler core, there are at least two
> >>>>>>> outside controllers (e.g., devfreq-cooling.c and throttler driver).
> >>>>>>> As I knew the problem about conflict, I cannot agree the temporary
> >>>>>>> method. OPP interface is mandatory for devfreq in order to control
> >>>>>>> the OPP (frequency/voltage). In this situation, we have to try to
> >>>>>>> find the method through OPP interface.
> >>>>>>
> >>>>>> What do you mean with "temporary method"?
> >>>>>>
> >>>>>> We can try to find a method through the OPP interface, but at this
> >>>>>> point I'm not convinced that it is technically necessary or even
> >>>>>> preferable.
> >>>>>>
> >>>>>> Another inconvenient of the OPP approach for both devfreq-cooling.c
> >>>>>> and the throttler is that they have to bother with disabling all OPPs
> >>>>>> above/below the max/min (they don't/shouldn't have to care), instead
> >>>>>> of just telling devfreq the max/min.
> >>>>>
> >>>>> And a more important one: both drivers now have to keep track which
> >>>>> OPPs they enabled/disabled previously, done are the days of a simple
> >>>>> dev_pm_opp_enable/disable() in devfreq_cooling. Certainly it is
> >>>>> possible and not very complex to implement, but is it really the
> >>>>> best/a good solution?
> >>>>
> >>>>
> >>>> As I replied them right before, Each outside driver has their own throttling
> >>>> policy to control OPP entries. They don't care the requirement of other
> >>>> driver and cannot know the requirement of other driver. devfreq core can only
> >>>> recognize them and then only consider enabled OPP entris without disabled OPP entries.
> >>>>
> >>>> For example1,
> >>>>        | devfreq-cooling| throttler
> >>>> ---------------------------------------
> >>>> 500Mhz | disabled       | disabled
> >>>> 400Mhz | disabled       | disabled
> >>>> 300Mhz |                | disabled
> >>>> 200Mhz |                |
> >>>> 100Mhz |                |
> >>>> => devfreq driver can use only 100/200Mhz
> >>>>
> >>>>
> >>>> For example2,
> >>>>        | devfreq-cooling| throttler
> >>>> ---------------------------------------
> >>>> 500Mhz | disabled       | disabled
> >>>> 400Mhz | disabled       |
> >>>> 300Mhz | disabled       |
> >>>> 200Mhz |                |
> >>>> 100Mhz |                |
> >>>> => devfreq driver can use only 100/200Mhz
> >>>>
> >>>>
> >>>> For example3,
> >>>>        | devfreq-cooling| throttler
> >>>> ---------------------------------------
> >>>> 500Mhz | disabled       | disabled
> >>>> 400Mhz |                |
> >>>> 300Mhz |                |
> >>>> 200Mhz |                | disabled
> >>>> 100Mhz |                | disabled
> >>>> => devfreq driver can use only 300/400Mhz
> >>>
> >>> These are all cases without conflicts, my concern is about this:
> >>>
> >>>>        | devfreq-cooling| throttler
> >>>> ---------------------------------------
> >>>> 500Mhz | disabled       |
> >>>> 400Mhz | disabled       |
> >>>> 300Mhz |                | disabled
> >>>> 200Mhz |                | disabled
> >>>> 100Mhz |                | disabled
> >>>> => devfreq driver can't use any frequency?
> >>
> >> There are no any enabled frequency. Because device driver
> >> (devfreq-cooling, throttler) disable all frequencies.
> >>
> >> Outside drivers(devfreq-cooling, throttler) can enable/disable
> >> specific OPP entries. As I already commented, each outside driver
> >> doesn't consider the policy of other device driver about OPP entries.
> > 
> > And wouldn't it be preferable to have an interface that tries to avoid
> > this situation in the first place and has a clear policy for conflict
> > resolution?
> > 
> >> OPP interface is independent on devfreq and just control OPP entries.
> >> After that, devfreq just consider the only enabled OPP entries.
> >>
> >>>
> >>> Actually my above comment wasn't about this case, but about the
> >>> added complexity in devfreq-cooling.c and the throttler:
> >>>
> >>> A bit simplified partition_enable_opps() currently does this:
> >>>
> >>> for_each_opp(opp) {
> >>>   if (opp->freq <= max)
> >>>      opp_enable(opp)
> >>>   else
> >>>      opp_disable(opp)
> >>> }
> >>>
> >>> With the OPP usage/disable count this doesn't work any longer. Now we
> >>> need to keep track of the enabled/disabled state of the OPP, something
> >>> like:
> >>>
> >>> dev_pm_opp_enable(opp) {
> >>>   if (opp->freq <= max) {
> >>>     if (opp->freq > prev_max)
> >>>       opp_enable(opp)
> >>>   } else {
> >>>     if (opp->freq < prev_max)
> >>>       opp_disable(opp)
> >>>   }
> >>> }
> >>>
> >>> And duplicate the same in the throttler (and other possible
> >>> drivers). Obviously it can be done, but is there really any gain
> >>> from it?
> >>>
> >>> Instead they just could do:
> >>>
> >>> devfreq_verify_within_limits(policy/freq_pair, 0, max_freq)
> 
> I have a new question about using devfreq_verify_within_limits().
> 
> For example,
> Driver A has following opp-table
> - 500Mhz
> - 400Mhz
> - 300Mhz
> - 200Mhz
> - 100Mhz
> 
> Basically, driver A has following init value:
> - policy->min is 100
> - policy->max is 500
> 
> Driver B, devfreq_verify_within_limits(200, 300)
> 	policy->min is 200
> 	policy->max is 300
> Driver C, devfreq_verify_within_limits(300, 400)
> 	policy->min is 300
> 	policy->max is 300
> Driver D, devfreq_verify_within_limits(400, 500)
> 	policy->min is 400
> 	policy->max is 400
> 
> In result, it looks like the requirement of Driver B are disappeared.
> is it the intention of devfreq_verify_within_limits()?

The requirements are conflicting, neither
devfreq_verify_within_limits() nor any other mechanism can completely
resolve that. Any conflict resolution method will violate the
requirements of at least one client. I don't claim the current method
is necessarily the best, it's just one that takes a predictable
action, I'm open to other suggestions.

> >>> without being concerned about implementation details of devfreq.
> >>>
> >>
> >> I don't think so.
> > 
> > What are you referring to, the change that I claim that will be needed
> > in partition_enable_opps() when OPPs have usage/disable counts? If so,
> > how do you avoid that the function doesn't enable/disable an OPP that
> > was already enabled/disabled in the previous iteration?
> 
> Just about changes of "dev_pm_opp_enable(opp)".
> 
> > 
> >> dev_pm_opp_enable()/dev_pm_opp_disable() have to consider only one
> >> OPP entry without any other OPP entry.
> > 
> > I agree with this :)
> > 
> >> dev_pm_opp_enable()/dev_pm_opp_disable() can never know the other
> >> OPP entries. After some driver(devfreq-cooling.c and throttler) 
> >> enable or disable specific OPP entries, the remaining OPP entry 
> >> with enabled state will be considered on devfreq driver.
> > 
> > Having multiple drivers (or even a single one) enable and disable
> > OPPs independently and at the time of their choosing sounds like a
> > recipe for race conditions.
> > 
> > What happens if e.g. the devfreq core calls
> > dev_pm_opp_find_freq_ceil/floor() and right after returning another
> > driver disables the OPP? devfreq uses the disabled OPP. Probably not a\
> 
> devfreq doesn't use the disabled OPP.
> 
> For example,
> 1. devfreq-cooling.c disable/enable some OPP and OPP send notification about OPP changes.
> 2. devfreq receives the notification (devfreq_notifier_call() is executed)
> 3. devfreq_notifier_call() try to find scaling_min_freq/scaling_max_freq

3.5 The throttler runs and disables scaling_max_freq

> 4. devfreq_notifier_call() executes update_devfreq() in order to apply the OPP changes.
> 5. devfreq can consider only enabled frequencies right after dev_pm_opp_disable/enable()

Depending on the locking requirements within the throttler it *might*
be possible to avoid this situation if the throttler held
devfreq->lock while manipulating the OPPs. Not a great option though,
since we'd be dealing with devfreq implementation details in the
throttler.

If I am not mistaken the same can happen nowadays with devfreq_cooling.c:

<thermal governor>
  mutex_lock(tz->lock)
  thermal_cdev_update
    mutex_lock(cdev->lock)
      devfreq_cooling_set_cur_state
        partition_enable_opps

The governor function can run at the same time as
devfreq_notifier_call() / update_devfreq() and unless I overlooked
something there is nothing preventing partition_enable_opps() from
disabling OPPs while the devfreq functions are executing.

> > big deal if disabling the OPP is only a semantic question, but I
> > imagine there can be worse scenarios. Currently the only user of
> > dev_pm_opp_disable() besides devfreq_cooling.c is imx6q-cpufreq.c, and
> > it is well behaved and only disables OPPs during probe().
> 
> imx6q-cpufreq.c used the dev_pm_opp_disable() before calling
> dev_pm_opp_init_cpufreq_table(). After registered
> cpufreq_register_driver(), imx6q-cpufreq.c doesn't use the
> dev_pm_opp_disable/enable(). It means that dev_pm_opp_disable() of
> imx6q-cpufreq.c doesn't affect the frequency choice of cpufreq on
> the runtime after registered cpufreq driver.
>
> On the other hand, devfreq_cooling.c use dev_pm_opp_disable/enable()
> on the runtime after registering devfreq driver. It affect the
> frequency choice of devfreq on the runtime.

Exactly, that was my point. devfreq_cooling.c is currently the only
driver that uses the interface at runtime. And I have doubts that
proper synchronization is in place.

> > I keep missing a clear answer to the question in which sense
> > manipulating the OPPs in devfreq_cooling.c is superior over narrowing
> > down the frequency during DEVFREQ_ADJUST, which would avoid potential
> > races and allow to resolve conflicts. Does it allow for some
> 
> You mentioned the race conditions eariler. Actually, I don't know
> the potential races.

Does my above analysis look correct to you, or did I miss some
mechanism that provides synchronization?

> > functionality that couldn't be achieved otherwise, does it make the
> > code significantly less complex, is some integration with the OPP
> > subsystem needed that I'm overlooking, is it more efficient, ...?
> > 
> > I'm not just insisting because I'm stubborn. I'd be happy to use any
> > interface that fits the bill, or to adjust one to fit the bill, but as
> > of now I mainly see drawbacks on the OPPs side and haven't seen
> > convincing arguments that it is really needed in devreq_cooling.c or a
> > better solution.
> 
> During we discussed, we knew that OPP doesn't provide all operation
> perfectly. As of now, OPP is standard framework to control the pair
> of frequency/voltage. devfreq need to use OPP to the pair of
> frequency/voltage. Even if OPP doesn't provide all of our
> requirements, I think that devfreq should use OPP interface after
> updating OPP framework, instead of adding other functionality to
> control frequency in outside driver(devfreq-cooling.c, throttler).

I think it's great if devfreq uses the OPP interface internally, and
disables OPPs *after* having received the requirements from all
outside drivers and resolved potential conflicts.

For narrowing down the frequency range from outside drivers - even
with the extended OPP interface we are discussing - there are multiple
caveats and so far not a single *technical* argument in it's favor
(and I'm really in favor of using standard interfaces when suitable
instead of implementing custom solutions, though in this case the
'custom solution' is a simple function call).

> I asked to Viresh (OPP maintainer). If dev_pm_opp_enable() and
> dev_pm_opp_disable() don't support the usecase on multiple device
> drivers, opp interface could not be used in order to support both
> devfreq-cooling.c and throttler. So, We better to wait the opinion
> from OPP maintainer.

Yes, it would be interesting to know Viresh's opinion, maybe I'm
completely wrong and there is an elegant solution using the
OPPs. Hopefully this thread didn't get buried in his inbox, it might
be worth to ping him separately.

Thanks

Matthias