Re: Alternatives to /sys/kernel/debug/wakeup_sources

[Tri Vo <trong@android.com>]

On Tue, Jun 18, 2019 at 2:23 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
>
> On Tuesday, June 18, 2019 10:17:16 PM CEST Sandeep Patil wrote:
> >
> > Hi Rafael, Viresh etc.
> >
> > On Tue, Jun 11, 2019 at 10:31:16AM -0700, Tri Vo wrote:
> > > On Tue, Jun 4, 2019 at 5:23 PM Tri Vo <trong@android.com> wrote:
> > > >
> > > > Hello Rafael,
> > > >
> > > > Currently, Android reads wakeup sources statistics from
> > > > /sys/kernel/debug/wakeup_sources in production environment. This
> > > > information is used, for example, to report which wake lock prevents
> > > > the device from suspending.
> >
> > Android's usage of the 'wakeup_sources' from debugfs can is linked at[1].
> > Basically, android's battery stats implementation to plot history for suspend
> > blocking wakeup sources over device's boot cycle. This is used both for power
> > specific bug reporting but also is one of the stats that will be used towards
> > attributing the battery consumption to specific processes over the period of
> > time.
> >
> > Android depended on the out-of-tree /proc/wakelocks before and now relies on
> > wakeup_sources debugfs entry heavily for the aforementioned use cases.
> >
> > > >
> > > > Android userspace reading wakeup_sources is not ideal because:
> > > > - Debugfs API is not stable, i.e. Android tools built on top of it are
> > > > not guaranteed to be backward/forward compatible.
> > > > - This file requires debugfs to be mounted, which itself is
> > > > undesirable for security reasons.
> > > >
> > > > To address these problems, we want to contribute a way to expose these
> > > > statistics that doesn't depend on debugfs.
> > > >
> > > > Some initial thoughts/questions: Should we expose the stats in sysfs?
> > > > Or maybe implement eBPF-based solution? What do you think?
> >
> > We are going through Android's out-of-tree kernel dependencies along with
> > userspace APIs that are not necessarily considered "stable and forever
> > supported" upstream. The debugfs dependencies showed up on our radar as a
> > result and so we are wondering if we should worry about changes in debugfs
> > interface and hence the question(s) below.
> >
> > So, can we rely on /d/wakeup_sources to be considered a userspace API and
> > hence maintained stable as we do for other /proc and /sys entries?
> >
> > If yes, then we will go ahead and add tests for this in LTP or
> > somewhere else suitable.
>
> No, debugfs is not ABI.
>
> > If no, then we would love to hear suggestions for any changes that need to be
> > made or we simply just move the debugfs entry into somewhere like
> > /sys/power/ ?
>
> No, moving that entire file from debugfs into sysfs is not an option either.
>
> The statistics for the wakeup sources associated with devices are already there
> under /sys/devices/.../power/ , but I guess you want all wakeup sources?
>
> That would require adding a kobject to struct wakeup_source and exposing
> all of the statistics as separate attributes under it.  In which case it would be
> good to replace the existing wakeup statistics under /sys/devices/.../power/
> with symbolic links to the attributes under the wakeup_source kobject.

Thanks for your input, Rafael! Your suggestion makes sense. I'll work
on a patch for this.
>
> > As a side effect, if the entry moves out of debugfs, Android can run without
> > mounting debugfs in production that I assume is a good thing.
>
> And really Android developers might have thought about this a bit earlier.

I'm still learning about kernel development. And Android has made
missteps before. So I figured it's a good idea to ask first :)

Thanks!