Re: [PATCH] hwmon (ina3221) Add single-shot mode support

[Guenter Roeck]

On Tue, Nov 13, 2018 at 04:11:42PM -0800, Nicolin Chen wrote:
> Hi Guenter,
> 
> On Tue, Nov 13, 2018 at 09:21:02AM -0800, Guenter Roeck wrote:
> > > > > INA3221 supports both continuous and single-shot modes. When
> > > > > running in the continuous mode, it keeps measuring the inputs
> > > > > and converting them to the data register even if there are no
> > > > > users reading the data out. In this use case, this could be a
> > > > > power waste.
> > > > > 
> > > > > So this patch adds a single-shot mode support so that ina3221
> > > > > could do measurement and conversion only if users trigger it,
> > > > > depending on the use case where it only needs to poll data in
> > > > > a lower frequency.
> > > > > 
> > > > > The change also exposes "mode" and "available_modes" nodes to
> > > > > allow users to switch between two operating modes.
> > > > > 
> > > > Lots and lots of complexity for little gain. Sorry, I don't see
> > > > the point of this change.
> > > 
> > > The chip is causing considerable power waste on battery-powered
> > > devices so we typically use it running in the single-shot mode.
> > 
> > And you need to be able to do that with a sysfs attribute ?
> > Are you planning to have some code switching back and forth
> > between the modes ?
> >
> > You'll need to provide a good rationale why this needs to be
> > runtime configurable.
> 
> Honestly, our old downstream driver didn't expose it via sysfs.
> Instead, it had a built-in "governor" to switch modes based on
> the CPU hotplug state and cpufreq. However, the interface used
> to register a CPU hotplug notification was already deprecated.
> And I don't feel this governor is generic enough to be present
> in the mainline code.
> 
> For me, it's not that necessary to be a sysfs attribute. I try
> to add it merely because I cannot find a good criteria for the
> mode switching in a hwmon driver. So having an open sysfs node
> may allow user space power daemon to decide its operating mode,
> since it knows which power mode the system is running at: full
> speed (charging/charged) or power saving (on-battery), and it
> knows how often this exact service will poll the sensor data.
> 
That is bad, because it is not a generic implementation.
Userspace would have to account for each individual driver
one by one.

> An alternative way (without the sysfs node), after looking at
> other hwmon code, could be to have a timed polling thread and
> read data using an update_interval value from ABI. This might
> turn out to be more complicated as it'll also involve settings
> of hardware averaging and conversion time. Above all, I cannot
> figure out a good threshold of update_interval to switch modes.
> 
update_interval should only be used if it can be configured
into hardware, not to trigger a polling thread. It should only
be used in the driver to determine caching intervals.

> If you can give some advice of a better implementation, that'd
> be great.
> 
>From your description, the only real feasible solution would be a
generic one, with a well defined interface either to userspace or
to the kernel. The best I can think of would be an in-kernel API
setting the power operational mode via callbacks. Alternative would
be a generic ability to set power operational mode from userspace,
using an ABI which applies to all drivers, not just one.

I don't know if any of those interfaces exists. If not, this will
require a discussion with upstream kernel maintainers, maybe with
a strawman proposal (set of patches). We can't just implement a
driver specific solution.

Guenter