Re: [PATCH] PM: domains: Ensure genpd_debugfs_dir exists before remove

[Hsin-Yi Wang]

On Tue, Jul 5, 2022 at 7:51 PM Greg Kroah-Hartman
<gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Tue, Jul 05, 2022 at 07:06:41PM +0800, Hsin-Yi Wang wrote:
> > On Tue, Jul 5, 2022 at 6:38 PM Greg Kroah-Hartman
> > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Jul 05, 2022 at 05:49:47PM +0800, Hsin-Yi Wang wrote:
> > > > genpd_debug_remove() may be indirectly called from others while
> > > > genpd_debugfs_dir is not yet set. Make sure genpd_debugfs_dir exists
> > > > before remove the sub components, otherwise components under
> > > > /sys/kernel/debug may be accidentally removed.
> > > >
> > > > Fixes: 718072ceb211 ("PM: domains: create debugfs nodes when adding power domains")
> > > > Signed-off-by: Hsin-Yi Wang <hsinyi@xxxxxxxxxxxx>
> > > > ---
> > > > An example:
> > > > scpsys_probe() in drivers/soc/mediatek/mtk-pm-domains.c indirectly calls
> > > > genpd_debug_remove() on probe fail, causing /sys/kernel/debug/usb to be
> > > > removed.
> > > > ---
> > > >  drivers/base/power/domain.c | 3 +++
> > > >  1 file changed, 3 insertions(+)
> > > >
> > > > diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
> > > > index 3e86772d5fac5..5a2e0232862e0 100644
> > > > --- a/drivers/base/power/domain.c
> > > > +++ b/drivers/base/power/domain.c
> > > > @@ -222,6 +222,9 @@ static void genpd_debug_remove(struct generic_pm_domain *genpd)
> > > >  {
> > > >       struct dentry *d;
> > > >
> > > > +     if (!genpd_debugfs_dir)
> > > > +             return;
> > > > +
> > > >       d = debugfs_lookup(genpd->name, genpd_debugfs_dir);
> > > >       debugfs_remove(d);
> > >
> > > Why not just change this to be:
> > >         debugfs_remove(debugfs_lookup(genpd->name, debugfs_lookup("pm_genpd", NULL)));
> > If pm_genpd hasn't been created yet,  debugfs_lookup("pm_genpd", NULL)
> > will return null.
>
> And how is this codepath being called if pm_genpd is not created yet?
> Surely you are not relying on the presence of a debugfs file to
> determine that?
>

Caller didn't directly call genpd_debug_remove(). The flow is as follows:

Normally, scpsys will create pm domain by:
scpsys_probe()
   --> scpsys_add_one_domain()
     --> pm_genpd_init()
       --> genpd_debug_add()


If something fails, it will do the cleanup:
scpsys_probe()
  --> scpsys_domain_cleanup()
     --> scpsys_remove_one_domain()
       --> pm_genpd_remove()
          --> genpd_remove()
            --> genpd_debug_remove()

genpd_debug_add() checks if genpd_debugfs_dir is init by a
late_initcall genpd_debug_init(). If it's NULL, it will return
directly without creating anything. Later when genpd_debug_init() is
called, it will call genpd_debug_add() again.

pm_genpd_remove() still needs to be called on the cleanup path to free
other stuff, but if genpd_debug_init() hasn't happened,
genpd_debug_remove() should be a no-op, or genpd_remove() shouldn't
call it. (We can move the check there, but adding in
genpd_debug_remove() is more similar to what genpd_debug_add()
currently is.)

> > If genpd->name also exists under root debugfs, it will still be
> > deleted unintentionally, since NULL represents root debugfs.
> > Eg. one of the genpd->name is "usb", which is supposed to be added as
> > /sys/kernel/debug/pm_genpd/usb later. But pm_genpd is not yet created,
> > /sys/kernel/debug/usb will be removed.
>
> Ah, that's a bad name to pick :)
>
> But still, don't paper over this problem, please solve the root issue of
> never relying on the creation of a debugfs file to determine functional
> logic.
>
Currently we can't remove genpd_debug_add() called in pm_genpd_init()
and genpd_debug_remove() called in genpd_remove(). If some power
domain is created after genpd_debug_init(), their genpd subdomain
won't be registered in debugfs. genpd_debug_remove() also has similar
issues, eg. failed domain not removed after genpd_debug_init() is called.

> thanks,
>
> greg k-h