Re: [PATCH] PM / wakeup: Register wakeup class kobj after device is added

[Stephen Boyd]

Quoting Tri Vo (2019-08-16 14:27:35)
> On Fri, Aug 16, 2019 at 7:56 AM Stephen Boyd <swboyd@xxxxxxxxxxxx> wrote:
> > diff --git a/drivers/base/power/sysfs.c b/drivers/base/power/sysfs.c
> > index 1b9c281cbe41..27ee00f50bd7 100644
> > --- a/drivers/base/power/sysfs.c
> > +++ b/drivers/base/power/sysfs.c
> > @@ -5,6 +5,7 @@
> >  #include <linux/export.h>
> >  #include <linux/pm_qos.h>
> >  #include <linux/pm_runtime.h>
> > +#include <linux/pm_wakeup.h>
> >  #include <linux/atomic.h>
> >  #include <linux/jiffies.h>
> >  #include "power.h"
> > @@ -661,14 +662,21 @@ int dpm_sysfs_add(struct device *dev)
> >                 if (rc)
> >                         goto err_runtime;
> >         }
> > +       if (dev->power.wakeup) {
> 
> This conditional checks for the situation when wakeup source
> registration have been previously attempted, but failed at
> wakeup_source_sysfs_add(). My concern is that it's not easy to
> understand what this check does without knowing exactly what
> device_wakeup_enable() does to dev->power.wakeup before we reach this
> point.

Oh, actually this is wrong. It should be a check for
dev->power.wakeup->dev being non-NULL. That's the variable that's set by
wakeup_source_sysfs_add() upon success. So I should make it:

	if (dev->power.wakeup && !dev->power.wakeup->dev)

And there's the problem that CONFIG_PM_SLEEP could be unset. Let me fix
it up with a new inline function like device_has_wakeup_dev().

> 
> > +               rc = wakeup_source_sysfs_add(dev, dev->power.wakeup);
> > +               if (rc)
> > +                       goto err_wakeup;
> > +       }
> >         if (dev->power.set_latency_tolerance) {
> >                 rc = sysfs_merge_group(&dev->kobj,
> >                                        &pm_qos_latency_tolerance_attr_group);
> >                 if (rc)
> > -                       goto err_wakeup;
> > +                       goto err_wakeup_source;
> >         }
> >         return 0;
> >
> > + err_wakeup_source:
> > +       wakeup_source_sysfs_remove(dev->power.wakeup);
> >   err_wakeup:
> >         sysfs_unmerge_group(&dev->kobj, &pm_wakeup_attr_group);
> >   err_runtime:
> > diff --git a/drivers/base/power/wakeup.c b/drivers/base/power/wakeup.c
> > index f7925820b5ca..5817b51d2b15 100644
> > --- a/drivers/base/power/wakeup.c
> > +++ b/drivers/base/power/wakeup.c
> > @@ -220,10 +220,12 @@ struct wakeup_source *wakeup_source_register(struct device *dev,
> >
> >         ws = wakeup_source_create(name);
> >         if (ws) {
> > -               ret = wakeup_source_sysfs_add(dev, ws);
> > -               if (ret) {
> > -                       wakeup_source_free(ws);
> > -                       return NULL;
> > +               if (!dev || device_is_registered(dev)) {
> 
> Is there a possible race condition here? If dev->power.wakeup check in
> dpm_sysfs_add() is done at the same time as device_is_registered(dev)
> check here, then wakeup_source_sysfs_add() won't ever be called?

The same race exists for device_set_wakeup_capable() so I didn't bother
to try to avoid it. I suppose wakeup_source_sysfs_add() could run
completely, allocate the device and set the name, etc., but not call
device_add() and then we can set ws->dev and call device_add() under a
mutex so that we keep a very small window where the wakeup class is
published to sysfs. Or just throw a big mutex around the whole wakeup
class creation path so that there isn't a chance of a race. But really,
is anyone going to call device_set_wakeup_*() on a device that is also
being added to the system? Seems unlikely.

> 
> > +                       ret = wakeup_source_sysfs_add(dev, ws);
> > +                       if (ret) {
> > +                               wakeup_source_free(ws);