PM: Limit race conditions between runtime PM and system sleep (v2)

One of the roles of the PM core is to prevent different PM callbacks
executed for the same device object from racing with each other.
Unfortunately, after commit e8665002
(PM: Allow pm_runtime_suspend() to succeed during system suspend)
runtime PM callbacks may be executed concurrently with system
suspend/resume callbacks for the same device.

The main reason for commit e8665002
was that some subsystems and device drivers wanted to use runtime PM
helpers, pm_runtime_suspend() and pm_runtime_put_sync() in
particular, for carrying out the suspend of devices in their
.suspend() callbacks.  However, as it's been determined recently,
there are multiple reasons not to do so, inlcuding:

 * The caller really doesn't control the runtime PM usage counters,
   because user space can access them through sysfs and effectively
   block runtime PM.  That means using pm_runtime_suspend() or
   pm_runtime_get_sync() to suspend devices during system suspend
   may or may not work.

 * If a driver calls pm_runtime_suspend() from its .suspend()
   callback, it causes the subsystem's .runtime_suspend() callback to
   be executed, which leads to the call sequence:

   subsys->suspend(dev)
      driver->suspend(dev)
         pm_runtime_suspend(dev)
            subsys->runtime_suspend(dev)

   recursive from the subsystem's point of view.  For some subsystems
   that may actually work (e.g. the platform bus type), but for some
   it will fail in a rather spectacular fashion (e.g. PCI).  In each
   case it means a layering violation.

 * Both the subsystem and the driver can provide .suspend_noirq()
   callbacks for system suspend that can do whatever the
   .runtime_suspend() callbacks do just fine, so it really isn't
   necessary to call pm_runtime_suspend() during system suspend.

 * The runtime PM's handling of wakeup devices is usually different
   from the system suspend's one, so .runtime_suspend() may simply be
   inappropriate for system suspend.

 * System suspend is supposed to work even if CONFIG_PM_RUNTIME is
   unset.

 * The runtime PM workqueue is frozen before system suspend, so if
   whatever the driver is going to do during system suspend depends
   on it, that simply won't work.

Still, there is a good reason to allow pm_runtime_resume() to
succeed during system suspend and resume (for instance, some
subsystems and device drivers may legitimately use it to ensure that
their devices are in full-power states before suspending them).
Moreover, there is no reason to prevent runtime PM callbacks from
being executed in parallel with the system suspend/resume .prepare()
and .complete() callbacks and the code removed by commit
e8665002 went too far in this
respect.  On the other hand, runtime PM callbacks, including
.runtime_resume(), must not be executed during system suspend's
"late" stage of suspending devices and during system resume's "early"
device resume stage.

Taking all of the above into consideration, make the PM core
acquire a runtime PM reference to every device and resume it if
there's a runtime PM resume request pending right before executing
the subsystem-level .suspend() callback for it.  Make the PM core
drop references to all devices right after executing the
subsystem-level .resume() callbacks for them.  Additionally,
make the PM core disable the runtime PM framework for all devices
during system suspend, after executing the subsystem-level .suspend()
callbacks for them, and enable the runtime PM framework for all
devices during system resume, right before executing the
subsystem-level .resume() callbacks for them.

Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Kevin Hilman <khilman@ti.com>