On Mon, May 21, 2018 at 08:48:32AM -0700, Tony Lindgren wrote:
> * Johan Hovold <johan@xxxxxxxxxx> [180521 13:50]:
> > On Thu, May 17, 2018 at 10:10:38AM -0700, Tony Lindgren wrote:
> > > * Johan Hovold <johan@xxxxxxxxxx> [180517 10:12]:
> > > Well in that case we should just stick with -1 value for the
> > > autosuspend. And just do pm_runtime_put_sync_suspend() after
> > > probe and on close.
> >
> > That won't work either as a negative autosuspend delay prevents runtime
> > suspend completely (by grabbing an extra RPM reference).
>
> Well so negative autosuspend delay is working as documented then :)
Yes, indeed. All too well. ;)
> > > > I fail to see how we can implement this using the current toolbox. What
> > > > you're after here is really a mechanism for selecting between two
> > > > different runtime PM schemes at runtime:
> > > >
> > > > 1. normal serial RPM, where the controller is active while the
> > > > port is open (this should be the safe default)
> > >
> > > Agreed. And that is the case already.
> >
> > Yes, but it's not really the case today as since omap-serial (and
> > 8250-omap) sets a negative autosuspend at probe and hence does not
> > runtime-suspend when the port is closed. So that's the long-standing bug
> > which needs fixing.
>
> Yes the bug for closed ports needs to be fixed for sure.
I did some forensic on this and it seems this problem has "always" been
there. Specifically, closed ports have never been runtime suspended
unless a non-negative autosuspend delay has been set by user space since
fcdca75728ac ("ARM: OMAP2+: UART: Add runtime pm support for omap-serial
driver") which was merged seven years ago.
So while it would certainly be nice to save some more power by default,
this would really be a new feature rather than a bug or regression fix
(which reduces the urgency for this issue somewhat too).
> > > > 2. aggressive serial RPM, where the controller is allowed to
> > > > suspend while the port is open even though this may result in
> > > > lost characters when waking up on incoming data
> > >
> > > In this case it seems that the only thing needed is to just
> > > configure the autosuspend delay for the parent port. The use of
> > > -1 has been around since the start of runtime PM AFAIK, so maybe
> > > we should just document it. I guess we could also introduce
> > > pm_runtime_block_autoidle_unless_configured() :)
> >
> > The implications of a negative autosuspend delay are already documented
> > (in Documentation/power/runtime_pm.txt); it's just the omap drivers that
> > gets it wrong when trying to do things which aren't currently supported
> > (and never have been).
> >
> > So I still think we need a new mechanism for this.
>
> Well if you have some better mechanism in mind let's try it out. Short of
> sprinkling pm_runtime_force_suspend/resume calls all over, I'm out of ideas
> right now.
Yeah, that would be too much of a hack and likely wouldn't work either
(and we really should do away with those _force calls altogether).
I've been thinking a bit too much about this already, but it may be
possible to use the pm QoS framework for this. A resume latency can be
set through sysfs where "n/a" is defined to mean "no latency accepted"
(i.e. controller remains always-on while port is open) and "0" means
"any latency accepted" (i.e. omap aggressive serial RPM is allowed).
Now, implementing this may get a little tricky as we want to be able to
change this setting on the fly (consider consoles) and we need to figure
out the interaction with serdev (user space should probably not be
allowed to request a resume latency for ports used by serdev).
I'd be happy to dig into this some more, but not in my spare time I'm
afraid.
> > > > For normal ttys, we need a user-space interface for selecting between
> > > > the two, and for serdev we may want a way to select the RPM scheme from
> > > > within the kernel.
> > > >
> > > > Note that with my serdev controller runtime PM patch, serdev core could
> > > > always opt for aggressive PM (as by default serdev core holds an RPM
> > > > reference for the controller while the port is open).
> > >
> > > So if your serdev controller was to set the parent autosuspend
> > > delay on open() and set it back on close() this should work?
> >
> > Is it really the job of a serdev driver to set the autosuspend delay of
> > a parent controller? Isn't this somethings which depends on the
> > characteristics of the controller (possibly configurable by user space)
> > such as the cost of runtime suspending and resuming?
>
> Only in some cases will the serdev driver know it's safe to configure
> the parent controller. Configuring the parent controller from userspace
> works just fine as we've seen for years now.
Yes, user space may override the default settings provided by the serial
driver, but a serdev driver, in contrast, knows nothing about the
underlying serial hardware.
> > The patch I posted works with what we have today; if a parent serial
> > controller driver uses aggressive runtime PM by default or after having
> > been configured through sysfs to do so.
>
> Yeah let's stick with configuring the parent controller from userspace
> for now at least.
Yep, status quo works for the time being (since this isn't a
regression).
> > What I'm getting at here is that the delay should be set by the serial
> > driver implementing aggressive runtime PM. Then all we need is a
> > mechanism to determine whether an extra RPM reference should be taken at
> > tty open or not (configurable by user space, defaulting to yes).
>
> OK yeah some additional on/off switch seems to be missing here.
As mentioned above, PM QoS resume latency may possibly be used, and
otherwise me may able to define a new (generic) QoS flag for this.
> > Specifically, the serial drivers themselves would always use
> > autosuspend and not have to deal with supporting the two RPM schemes
> > (normal vs aggressive runtime PM).
>
> OK. So if I understand your idea right, we could have autosuspend timeout
> set to 3000ms in the 8250_omap.c but still default to RPM blocked?
> Then user can enable aggressive PM via /sys as desired, right?
Not RPM blocked; the ports must always be able to suspend when the port
is closed. But user space should be able to enable the aggressive
(active) runtime PM via sysfs independently of the autosuspend delay,
yes.
Thanks,
Johan
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
