Re: [RFC][PATCH 2/3] PM / sleep: Mechanism to avoid resuming runtime-suspended devices unnecessarily

[Jacob Pan]

On Fri, 16 May 2014 23:08:01 +0200
"Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx> wrote:

> On Friday, May 16, 2014 08:20:55 AM Jacob Pan wrote:
> > On Thu, 15 May 2014 11:58:55 -0400 (EDT)
> > Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
> > 
> > > On Thu, 15 May 2014, Jacob Pan wrote:
> > > 
> > > > On Thu, 15 May 2014 10:29:42 -0400 (EDT)
> > > > Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
> > > > 
> > > > > On Thu, 15 May 2014, Jacob Pan wrote:
> > > > > 
> > > > > > > > should we respect ignore_children flag here? not all
> > > > > > > > parent devices create children with proper .prepare()
> > > > > > > > function. this allows parents override children.
> > > > > > > > I am looking at USB, a USB device could have logical
> > > > > > > > children such as ep_xx, they don't go through the same
> > > > > > > > subsystem .prepare().
> > > > > > > 
> > > > > > > Well, I'm not sure about that.  Let me consider that for a
> > > > > > > while.
> > > > > > OK. let me be more clear about the situation i see in USB.
> > > > > > Correct me if I am wrong, a USB device will always has at
> > > > > > least one endpoint/ep_00 as a kid for control pipe, it is a
> > > > > > logical device. So when device_prepare() is called, its
> > > > > > call back is NULL which makes .direct_complete = 0. Since
> > > > > > children device suspend is called before parents, the
> > > > > > parents .direct_complete flag will always get cleared.
> > > > > > 
> > > > > > What i am trying to achieve here is to see if we avoid
> > > > > > resuming built-in (hardwired connect_type) non-hub USB
> > > > > > devices based on this new patchset. E.g. we don't want to
> > > > > > resume/suspend USB camera every time in system
> > > > > > suspend/resume cycle if they are already rpm suspended. We
> > > > > > can save ~100ms resume time for the devices we have tested.
> > > > > 
> > > > > This is a good point, but I don't think it is at all related
> > > > > to ignore_children.
> > > > > 
> > > > > Instead, it seems that the best way to solve it would be to
> > > > > add a ->prepare() handler for usb_ep_device_type that would
> > > > > always turn on direct_complete.
> > > > > 
> > > > yeah, that would solve the problem with EP device type. But what
> > > > about other subdevices. e.g. for USB camera, uvcvideo device? We
> > > > can add .prepare(return 1;) for each level but would it be
> > > > better to have a flag similar to ignore_children if not
> > > > ignore_children itself.
> > > 
> > > Something like that could always be added.
> > or, how about if a device's .prepare() is NULL, we could
> > assume .direct_resume() should be set. i.e.
> 
> You mean direct_complete (which is a flag, not a function), I suppose?
> 
yes.
> Wouldn't that go a bit too far?  It seems to be based on the
> assumption that all devices having no ->prepare() callback can be
> safely left in runtime suspend over a system suspend/resume cycle,
> but is that assumption actually satisfied for all such devices?
> 
yes, I agree it is risky though i don't see problems with my limited
testing. But on the other side, it is too strict.
I also tried adding .prepare( return 1;) to usb_ep_device_type pm ops,
that didn't work either. The reason is that ep devices don't support
runtime pm (disable_depth > 0). I think in this case ignore_children
flag should be the right indicator to ignore pm_runtime_suspended()?

> > --- a/drivers/base/power/main.c
> > +++ b/drivers/base/power/main.c
> > @@ -1539,7 +1539,7 @@ static int device_prepare(struct device *dev,
> > pm_message_t state) pm_runtime_put(dev);
> >                 return ret;
> >         }
> > -       dev->power.direct_complete = ret > 0 && state.event ==
> > PM_EVENT_SUSPEND
> > +       dev->power.direct_complete = (!callback || ret > 0) &&
> > state.event == PM_EVENT_SUSPEND && pm_runtime_suspended(dev);
> >         dev_dbg(dev, "%s:direct_complete %d, info %s\n", __func__,
> > dev->power.direct_complete, info);
> > 
> > > 
> > > > Actually, I don't understand why this is not related to
> > > > ignore_children. Could you explain?
> > > 
> > > It's hard to explain why two things are totally separate.  Much
> > > better for you to describe why you think they _are_ related, so
> > > that I can explain how you are wrong.
> > > 
> > > > If the parent knows it can ignore children and already rpm
> > > > suspended, why do we still ask children?
> > > 
> > > The "ignore_children" flag doesn't mean that the parent can ignore
> > > its children.  It means that the PM core is allowed to do a
> > > runtime suspend of the parent while leaving the children at full
> > > power.
> > > 
> > > In particular, it doesn't mean that the children's ->suspend()
> > > callback will work correctly if it is called while the parent is
> > > runtime suspended.
> > that explains my question about ignore_chilren flag. thanks.
> > > 
> > > Alan Stern
> > > 
> > 
> > [Jacob Pan]
> > --
> > To unsubscribe from this list: send the line "unsubscribe
> > linux-acpi" in the body of a message to majordomo@xxxxxxxxxxxxxxx
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

[Jacob Pan]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/