Re: [RFC][PATCH 1/3] PM / sleep: Flag to speed up suspend-resume of runtime-suspended devices

[Rafael J. Wysocki]

On Sunday, May 11, 2014 12:46:10 PM Alan Stern wrote:
> On Sat, 10 May 2014, Rafael J. Wysocki wrote:
> 
> > On Thursday, May 08, 2014 09:52:18 PM Alan Stern wrote:
> > > On Thu, 8 May 2014, Rafael J. Wysocki wrote:
> > > 
> > > > Well, no.
> > > > 
> > > > The reason why that doesn't work is because ->prepare() callbacks are
> > > > executed in the reverse order, so the perent's ones will be run before
> > > > the ->prepare() of the children.  Thus if ->prepare() sets the flag
> > > > with the expectation that ->suspend() (and the subsequent callbacks)
> > > > won't be executed, that expectation may not be met actually.
> > > 
> > > That's true also if the flag gets set in ->suspend(), isn't it?  A
> > > driver may set direct_resume in its ->suspend() callback, expecting
> > > that the subsequent callbacks won't be executed.  But if a descendant
> > > hasn't also set its flag then the callbacks _will_ be executed.
> > 
> > No, that's not possible with the current patch, because __device_suspend() is
> > executed for descendants first and then for ancestors and it clears
> > direct_suspend for the parents of devices that don't have it set.  This means
> > that the ancestor's ->suspend() will see the flag clear if it is unset for
> > any of its descendants.
> > 
> > IOW, the only case in which the ancestor's ->suspend() sees the flag set is
> > when it has been set for all of its descendants.  Thus, if it leaves the
> > flag set, the late/early and noirq callbacks won't be executed for it.
> > 
> > Now, there is a reason for concern in that, because ->suspend() may set the
> > flag as a result of an error and that may lead to unexpected consequences.
> 
> Ah, my mistake; I should have read the patch more carefully.  I didn't
> realize that your plan was for subsystems/drivers to set the flag
> during ->prepare() and then clear it (or leave it set) during
> ->suspend().
> 
> > Then the parent will have direct_resume unset.  That is not a concern.
> > The only concern to me is possible errors in ->suspend() setting the
> > flag when it shouldn't.
> 
> So now one question is: Why would a subsystem or driver want to clear a
> flag that it had set earlier?  I can't think of any good reasons.  The
> only obvious possibility would be if the wakeup requirements got
> changed between ->prepare() and ->suspend(), but that should never
> happen because wakeup settings are changed by userspace and userspace
> will be frozen.
> 
> Another question is: Does a subsystem or driver need to know if the
> original flag setting couldn't be honored?  Again, I don't think it is 
> necessary to call ->suspend() just for this reason.  More precisely, I 
> think it will be good enough to call ->suspend() when the flag is clear 
> (either because a descendant device didn't set its flag or because the 
> device is no longer in runtime suspend); if the flag is set then there 
> is no reason to call ->suspend().  The subsystem can assume that 
> ->suspend() won't be called; then if it does get called, the subsystem 
> will realize something has changed.
> 
> Thus, a suitable algorithm now appears to be:
> 
> 	Have subsystems/drivers set the flag during ->prepare().  They
> 	don't even have to check if the device is runtime-suspended;
> 	if it isn't then the PM core will turn off the flag later.
> 
> 	In __device_suspend(), before invoking the ->supend() callback, 
> 	check the flag.  If it is still set and if the device is
> 	runtime-suspended (a barrier may be necessary here), skip 
> 	->suspend() and the following callbacks.  Otherwise clear the
> 	parent's flag and proceed as usual.
> 
> > > Several of these questions are a lot easier to answer if the flag gets 
> > > set during ->prepare() rather than ->suspend().

I actually decided to go that way with one difference.  I think it's better
to make the PM core own the new flag, so that bus types/drivers don't have
to set/clear it, so I got back to my very first idea about possibly returning
positive values from ->prepare().

The idea is this:

 - If ->prepare() returns a positive number, that means "this device is
   runtime-suspended and you can leave it like that if you do the same
   thing for all of its descendants".

 - If that happens, the PM core sets the new flag for the device in
   question *if* the device is indeed runtime-suspended *and* *if*
   the transition is a suspend (and not hibernation, for example).
   Otherwise, it clears the flag for the device.  All of that happens in
   device_prepare().

 - In __device_suspend() the PM core clears the flag for the device's
   parent if it is clear for the device to ensure that the flag will only
   be set for a device if it is also set for all of its descendants.

 - PM core skips ->suspend/late/noirq and ->resume/early/noirq for all devices
   having the flag set - so the flag can be called "direct_complete" as it
   causes the PM core to go directy for the ->complete() callback when set.

 - The ->complete() callback has to check direct_complete if ->prepare()
   returned a positive number previously and is responsible for further
   handling of the device.

> > I agree with that, but I have one concern about this approach.  Namely,
> > in that case the PM core has to use pm_runtime_resume() or equivalent to
> > resume devices with the flag set during the device resume stage.  Now,
> > in the next step we may want to leave certain devices suspended at that
> > point and the PM core has no way to tell which ones.  Also subsystems
> > don't really have a chance to tell it about that (they would need to
> > know in advance during ->prepare(), which is kind of unrealistic, or
> > perhaps it isn't).
> > 
> > However, if ->resume() is called for devices with the flag set, like in
> > my most recent patch, the subsystem may decide not to resume the device
> > if it knows enough about it.
> > 
> > This pretty much is my only concern here, so I'm open to ideas how to deal
> > with leaving devices suspended (if possible) during the device resume stage. :-)
> 
> This is a good question.  I'm not sure of the best answer at the 
> moment.
> 
> > For one, postponing the resume to ->complete() is an option, but it will have
> > to be done with care, because the ->complete() callbacks are executed
> > sequentially, so calling pm_runtime_resume() from there is rather out of the
> > question.
> 
> Calling pm_request_resume() would be okay, though.
> 
> There's another aspect to this we need to consider: hibernation.  I'm
> quite sure we don't want to come out of hibernation thinking that
> devices are still in their runtime-suspended states.  Skipping the
> callbacks for freeze and thaw would be all right in principle, and
> maybe even for poweroff, but not for restore.  And of course, during
> the restore stages, the last thing subsystems and drivers will remember
> happening is freeze -- which might mean you shouldn't skip the freeze
> callbacks either.

I agree.  I think the outline above should address this too.

I'm going to send a new version of the patches implementing the idea above.

Rafael

--
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/