Re: [RFT][PATCH 0/3] cpufreq / PM: QoS: Introduce frequency QoS and use it in cpufreq

[Leonard Crestez <leonard.crestez@nxp.com>]

On 2019-10-23 11:54 AM, Rafael J. Wysocki wrote:
> On Wed, Oct 23, 2019 at 4:20 AM Leonard Crestez <leonard.crestez@nxp.com> wrote:
>> On 2019-10-23 1:48 AM, Rafael J. Wysocki wrote:
>>> On Wed, Oct 23, 2019 at 12:06 AM Leonard Crestez
>>> <leonard.crestez@nxp.com> wrote:
>>>> I've been working on a series which add DEV_PM_QOS support to devfreq,
>>>> now at v9:
>>>>
>>>> Your third patch removes DEV_PM_QOS_FREQUENCY_MIN/MAX that my series
>>>> depends upon. I found the email on patchwork, hopefully the in-reply-to
>>>> header is OK?
>>>>
>>>> As far as I can tell the replacement ("frequency qos") needs constraints
>>>> to be managed outside the device infrastructure and it's not obviously
>>>> usable a generic mechanism for making "min_freq/max_freq" requests to a
>>>> specific device.
>>>
>>> You can add a struct freq_constrants pointer to struct dev_pm_info and
>>> use it just fine.  It doesn't have to be bolted into struct
>>> dev_pm_qos.
>>
>> I'm not sure what you mean by this? min/max_freq was already available
>> in dev_pm_qos so it's not clear why it would be moved somewhere else.
>> What I'm looking for is a mechanism to make min/max_freq requests on a
>> per-device basis and DEV_PM_QOS_MIN_FREQUENCY already did that.
>>
>> Reuse is good, right?
> 
> But they go away in patch 3 of this series as there are no users in
> the tree.  Sorry about that. >
>>>> I've read a bit through your emails and it seems the problem is that
>>>> you're dealing with dev_pm_qos on per-policy basis but each "struct
>>>> cpufreq_policy" can cover multiple CPU devices.
>>>>
>>>> An alternative solution which follows dev_pm_qos would be to add
>>>> notifiers for each CPU inside cpufreq_online and cpufreq_offline. This
>>>> makes quite a bit of sense because each CPU is a separate "device" with
>>>> a possibly distinct list of qos requests.
>>>
>>> But combining the lists of requests for all the CPUs in a policy
>>> defeats the idea of automatic aggregation of requests which really is
>>> what PM QoS is about.
>>
>> My primary interest is the "dev" part of dev_pm_qos: making pm_qos
>> requests tied to a specific device.
> 
> The list of requests needs to be associated with the user of the
> effective constraint.  If that is the device, it is all good.
The phrase "user of the effective constraint" is somewhat unclear.

I'm using the target device as dev for dev_pm_qos, not the requestor. 
This is consistent with how it was used for cpufreq: thermal called a 
dev_pm_qos_add_request on with dev = cpu_dev not a thermal sensor or 
anything else.

However looking at other dev_pm_qos users there are instances of a 
driver calling dev_pm_qos_add_request on it's own device but this is not 
a strict requirement, correct?

>>> There have to be two lists of requests per policy, one for the max and
>>> one for the min frequency >
>>>> If cpufreq needs a group of CPUs to run at the same frequency then it
>>>> should deal with this by doing dev_pm_qos_read_frequency on each CPU
>>>> device and picking a frequency that attempts to satisfy all constraints.
>>>
>>> No, that would be combining the requests by hand.
>>
>> It's just a loop though.
> 
> Yes, it is, and needs to be run on every change of an effective
> constraint for any CPU even if the total effective constraint doesn't
> change.  And, of course, the per-policy user space limits would need
> to be combined with that by hand.
> 
> Not particularly straightforward if you asked me.

Well, this cpu-to-policy aggregation could also use a pm_qos_constraint 
object instead of looping.

> Not to mention the fact that, say, cpu_cooling, has a per-policy list
> of requests anyway.
> 
>>>> Handling sysfs min/max_freq through dev_pm_qos would be of dubious
>>>> value, though I guess you could register identical requests for each CPU.
>>>>
>>>> I'm not familiar with what you're trying to accomplish with PM_QOS other
>>>> than replace the sysfs min_freq/max_freq files:
>>>
>>> QoS-based management of the frequency limits is not really needed for
>>> that.  The real motivation for adding it were things like thermal and
>>> platform firmware induced limits that all have their own values to
>>> combine with the ones provided by user space.
>>
>> Current users seem to be thermal-related. Do you care about min/max_freq
>> requests from stuff not directly tied to a CPU?
> 
> Yes, I do.
> 
> And they will need to add requests per policy.
> 
>>>> What I want to do is add
>>>> a driver using the interconnect driver which translates requests for
>>>> "bandwidth-on-a-path" into "frequency-on-a-device". More specifically a
>>>> display driver could request bandwidth to RAM and this would be
>>>> translated into min frequency for NoC and the DDR controller, both of
>>>> which implement scaling via devfreq:
>>>>
>>>> This is part of an effort to upstream an out-of-tree "busfreq" feature
>>>> which allows device device to make "min frequency requests" through an
>>>> entirely out-of-tree mechanism. It would also allow finer-grained
>>>> scaling that what IMX tree currently support.
>>>>
>>>> If you're making cpufreq qos constrains be "per-cpufreq-policy" then
>>>> it's not clear how you would handle in-kernel constraints from other
>>>> subsystems. Would users have to get a pointer to struct cpufreq_policy
>>>> and struct freq_constraints?
>>>
>>> Yes.
>>>
>>>> That would make object lifetime a nightmare!
>>>
>>> Why really?  It is not much different from the device PM QoS case
>>   >> Actually,  is a simple
>>> one-for-one replacement of the former.  As it turns out, all of its
>>> users have access to a policy object anyway already.
>>
>> All current users are very closely tied to cpufreq, what I had in mind
>> is requests from unrelated subsystems.
> 
> You can use cpufreq policy notifiers for that.  Add a request for each
> CPU in the policy (or for each related CPU if that is needed) to
> policy->constraints on CREATE_POLICY and remove them on REMOVE_POLICY.
> That's all you need to do.
> 
> BTW, the original code from Viresh did that already, I haven't changed
> it.  And it didn't have per-CPU lists of frequency requests for that
> matter, it used the ones in policy->cpu as the per-policy lists, which
> doesn't work.
> 
>> Browsing through the cpufreq core it seems that it's possible for a
>> struct cpufreq_policy to be created and destroyed at various points, the
>> simplest example being rmmod/modprobe on a cpufreq driver.
>>
>> The freq_qos_add_request function grabs a pointer to struct
>> freq_constraints, this can become invalid when cpufreq_policy is freed.
>>
>> I guess all users need to register a CPUFREQ_POLICY_NOTIFIER and make
>> sure to freq_qos_add_request every time?
> 
> Yes.
> 
> The policy is the user of the effective constraint anyway and holding
> on to a list of requests without a user of the effective constraint
> would be, well, not useful.
> 
>> Looking at your [PATCH 2/3] I  can't spot any obvious issue, thermal clamping
>> code seems to get the appropriate callbacks.
>>
>>>> But dev_pm_qos solves this by tying to struct device.
>>
>> The lifetime of "struct device" is already controlled by
>> get_device/put_device.
> 
> And why does this matter here?

My point is that dev_pm_qos is easier for consumers to use than dealing 
with cpufreq_policy lifetime and has less exposure to cpufreq 
implementation details.

But all current consumers seem to be appropriately coupled into cpufreq.

>>> Well, the cpufreq sysfs is per-policy and not per-CPU and we really
>>> need a per-policy min and max frequency in cpufreq, for governors etc.
>>
>> Aggregation could be performed at two levels:
>>
>> 1) Per cpu device (by dev_pm_qos)
>> 2) Per policy (inside cpufreq)
>>
>> The per-cpu dev_pm_qos notifier would just update a per-policy
>> pm_qos_constraints object. The second step could even be done strictly
>> inside the cpufreq core using existing pm_qos, no need to invent new
>> frameworks.
>>
>> Maybe dev_pm_qos is not a very good fit for cpufreq because of these
>> "cpu device versus cpufreq_policy" issues but it makes a ton of sense
>> for devfreq. Can you maybe hold PATCH 3 from this series pending further
>> discussion?
> 
> It can be reverted at any time if need be and in 5.4 that would be dead code.

I guess I can post v10 of my "devfreq pm qos" which starts by reverting 
"PATCH 3" of this series?

--
Regards,
Leonard