Re: [PATCH v2] thermal: add sysfs_notify on some attributes

From: Srikar Srimath Tirumala
Date: Tue Mar 29 2016 - 22:53:27 EST

On 03/28/2016 06:35 PM, Eduardo Valentin wrote:
On Tue, Mar 15, 2016 at 11:08:00PM +0000, Pandruvada, Srinivas wrote:
On Mon, 2016-03-14 at 11:12 -0700, Srikar Srimath Tirumala wrote:
Add a sysfs_notify on thermal_zone*/temp and cooling_device*/
cur_state whenever any trip is triggered or cur state is changed.

This change allows usermode apps to register themselves to get
notified, when certain thermal conditions occur and reduce their
workload. This workload throttling allows usermode to react before
hardware clocks are throttled and keep some critical apps running
reliably longer.
I think we need a combination of proposal in and this.

For example this patch notifies that some trip is violated, but that is
not enough for user space application to take any action. Some trips
violations user space may not care as this may be a transient one. The
patch from Eduardo address that by providing trip, temperature and last
temperature information. But that patch only address hot trips. I
understand why Eduardo doesn't want to be notified for passive trips as
there will be too many.

Yeah, my original intention was to avoid flooding userland, specially
through a pipe like the sysfs netlink, which is in use to deal with
other stuff.

How about if the number of notifications were to be reduced like in v3 of this patch? Or is the consensus not to use sysfs_notify or uevents at all?

So IMO we need some mechanism to turn off notification and decide what
notification will result in user space notifications.
On some x86 systems we have 10+ passive/active trips, this will results
in too many notifications. We may be in thermally sensitive zone, where
more code excecution is more heat.

Exactly, that has direct impact not only on the process waiting for
thermal notifications, but also on other process listening to sysfs.

We may have some mask of trips for which will result in notifications.
By default no notifications, unless some user space requests.

The complexity of such communication (or even the current status of
sysfs ABI) starts to reach limit of such channel. We may definitely
consider other means, such as /dev interface, just like IIO does.

I agree with your point about the overhead of generating these events and thought maybe the problem could be partially addressed by reducing the number of notifications being sent. So with that in mind I created a V3 of the patches which does this. Please do take a look at it when you get a chance.

During last LPC we discussed about using IIO for temperature threshold
notifications and I submitted multiple changes for that. Looks like we
also care of trip point changes. So I think we need more comprehensive
mechanism to address this.

May be we should have thermal mini summit during LPC again and decide a
comprehensive plan to address all asynchronous thermal notifications.

I have created a wiki for LPC 2016

Overall I believe we need to solve the (temperature) sensing in a more
structured way within the kernel. We have three subsystem that allow
performing temperature sensing. They are different in design and
concept, but still solve similar problems.

I went through the LPC presentation/patches and had a couple of questions.

The thermal iio device seems perfectly suited for sending large amounts of thermal data from the kernel framework to user space. Are there a lot of platforms where thermal throttling happens in user space? If so, Is this an indication that passive thermal throttling is going into the user space some time in the future?

Our use cases don't really require too many notifications. But there seem to be other use cases out there that require lot more thermal info from the kernel. Just trying to understand what those are and what needs to be done to arrive at a solution that benefits all.

I would still prefer to get this better architectured.

Of course, we do not need to wait until LPC to start drafting this.

Again, please lets generate enough quorum to run the micro conf.


Eduardo Valentin

Much appreciate your comments and inputs.
Best Regards