Re: [PATCH V4] thermal/core/power_allocator: reset thermal governor when trip point is changed

From: Lukasz Luba
Date: Fri Jun 23 2023 - 03:49:25 EST




On 6/22/23 19:27, Rafael J. Wysocki wrote:
On Tue, Jun 20, 2023 at 1:56 PM Lukasz Luba <lukasz.luba@xxxxxxx> wrote:



On 6/20/23 11:39, Rafael J. Wysocki wrote:
On Tue, Jun 20, 2023 at 12:19 PM Lukasz Luba <lukasz.luba@xxxxxxx> wrote:

Hi Rafael,


On 6/20/23 11:07, Rafael J. Wysocki wrote:
On Tue, Jun 20, 2023 at 11:46 AM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:

[snip]


Because this is up to the governor, the core has no clue what to do
with the return value from ->reset() and so there should be none.

As I said, governors can print whatever diagnostic messages they like
in that callback, but returning anything from it to the core is just
not useful IMV.

For the rest of the governors - it's up to them what they
report in case non-passive trip is updated...

Sure.


What Di is facing is in the issue under the bucket of
'handle_non_critical_trips()' when the governor just tries to
work on stale data - old trip temp.

Well, fair enough, but what about the other governors? Is this
problem limited to power_allocator?

IIUC the core fwk code - non of the governors would be needed
to handle the critical/hot trips. For the rest of the trip types
I would say it's up to the governor. In our IPA case this stale
data is used for power budget estimation - quite fundamental
step. Therefore, the reset and start from scratch would make more
sense.

I think other governors don't try to 'estimate' such
abstract power headroom based on temperature - so probably
they don't have to even implement the 'reset()' callback
(I don't know their logic that well).

So there seems to be a claim that IPA is the only governor needing the
->reset() callback, but I have not seen any solid analysis confirming
that. It very well may be the case, but then the changelog should
clearly explain why this is the case IMO.

I agree, the patch header doesn't explain that properly. Here is the
explanation for this Intelligent Power Allocator (IPA):

The IPA controls temperature using PID mechanism. It's a closed
feedback loop. That algorithm can 'learn' from the 'observed'
in the past reaction for it's control decisions and accumulates that
information in the part called 'error integral'. Those accumulated
'error' gaps are the differences between the set target value and the
actually achieved value. In our case the target value is the target
temperature which is coming from the trip point. That part is then used
with the 'I' (of PID) component, so we can compensate for those
'learned' mistakes.
Now, when you change the target temperature value - all your previous
learned errors won't help you. That's why Intelligent Power Allocator
should reset previously accumulated history.



For the 2nd case IIUC the code, we pass the 'trip.temperature'
and should be ready for what you said (modification of that value).

Generally speaking, it needs to be prepared for a simultaneous change
of multiple trip points (including active), in which case it may not
be useful to invoke the ->reset() callback for each of them
individually.

Although, that looks more cleaner IMO. Resetting one by one in
a temperature order would be easily maintainable, won't be?

I wouldn't call it maintainable really.

First of all, the trips may not be ordered. There are no guarantees
whatsoever that they will be ordered, so the caller may have to
determine the temperature order in the first place. This would be an
extra requirement that currently is not there.

Apart from this, I don't see any fundamental difference between the
case when trip points are updated via sysfs and when they are updated
by the driver. The governor should reset itself in any of those cases
and even if one trip point changes, the temperature order of all of
them may change, so the governor reset mechanism should be able to
handle the case when multiple trip points are updated at the same
time. While you may argue that this is theoretical, the ACPI spec
clearly states that this is allowed to happen, for example.

If you want a generic reset callback for governors, that's fine, but
make it generic and not specific to a particular use case.

I think we agree here, but probably having slightly different
implementation in mind. Based on you explanation I think you
want simply this API:
void (*reset)(struct thermal_zone_device *tz);

1. no return value
2. no specific trip ID as argument

Do you agree?
IMO such implementation and API would also work for this IPA
purpose. Would that work for the ACPI use case as well?