Re: [PATCH v2 07/22] drm/msm: Do rpm get sooner in the submit path

[Rob Clark]

On Mon, Nov 2, 2020 at 9:47 PM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
>
> On 27-10-20, 17:05, Viresh Kumar wrote:
> > It isn't that straight forward unfortunately, we need to make sure the
> > table doesn't get allocated for the same device twice, so
> > find+allocate needs to happen within a locked region.
> >
> > I have taken, not so straight forward, approach to fixing this issue,
> > lets see if this fixes it or not.
> >
> > -------------------------8<-------------------------
> >
> > diff --git a/drivers/opp/core.c b/drivers/opp/core.c
> > index 4ac4e7ce6b8b..6f4a73a6391f 100644
> > --- a/drivers/opp/core.c
> > +++ b/drivers/opp/core.c
> > @@ -29,6 +29,8 @@
> >  LIST_HEAD(opp_tables);
> >  /* Lock to allow exclusive modification to the device and opp lists */
> >  DEFINE_MUTEX(opp_table_lock);
> > +/* Flag indicating that opp_tables list is being updated at the moment */
> > +static bool opp_tables_busy;
> >
> >  static struct opp_device *_find_opp_dev(const struct device *dev,
> >                                       struct opp_table *opp_table)
> > @@ -1036,8 +1038,8 @@ static void _remove_opp_dev(struct opp_device *opp_dev,
> >       kfree(opp_dev);
> >  }
> >
> > -static struct opp_device *_add_opp_dev_unlocked(const struct device *dev,
> > -                                             struct opp_table *opp_table)
> > +struct opp_device *_add_opp_dev(const struct device *dev,
> > +                             struct opp_table *opp_table)
> >  {
> >       struct opp_device *opp_dev;
> >
> > @@ -1048,7 +1050,9 @@ static struct opp_device *_add_opp_dev_unlocked(const struct device *dev,
> >       /* Initialize opp-dev */
> >       opp_dev->dev = dev;
> >
> > +     mutex_lock(&opp_table->lock);
> >       list_add(&opp_dev->node, &opp_table->dev_list);
> > +     mutex_unlock(&opp_table->lock);
> >
> >       /* Create debugfs entries for the opp_table */
> >       opp_debug_register(opp_dev, opp_table);
> > @@ -1056,18 +1060,6 @@ static struct opp_device *_add_opp_dev_unlocked(const struct device *dev,
> >       return opp_dev;
> >  }
> >
> > -struct opp_device *_add_opp_dev(const struct device *dev,
> > -                             struct opp_table *opp_table)
> > -{
> > -     struct opp_device *opp_dev;
> > -
> > -     mutex_lock(&opp_table->lock);
> > -     opp_dev = _add_opp_dev_unlocked(dev, opp_table);
> > -     mutex_unlock(&opp_table->lock);
> > -
> > -     return opp_dev;
> > -}
> > -
> >  static struct opp_table *_allocate_opp_table(struct device *dev, int index)
> >  {
> >       struct opp_table *opp_table;
> > @@ -1121,8 +1113,6 @@ static struct opp_table *_allocate_opp_table(struct device *dev, int index)
> >       INIT_LIST_HEAD(&opp_table->opp_list);
> >       kref_init(&opp_table->kref);
> >
> > -     /* Secure the device table modification */
> > -     list_add(&opp_table->node, &opp_tables);
> >       return opp_table;
> >
> >  err:
> > @@ -1135,27 +1125,64 @@ void _get_opp_table_kref(struct opp_table *opp_table)
> >       kref_get(&opp_table->kref);
> >  }
> >
> > +/*
> > + * We need to make sure that the OPP table for a device doesn't get added twice,
> > + * if this routine gets called in parallel with the same device pointer.
> > + *
> > + * The simplest way to enforce that is to perform everything (find existing
> > + * table and if not found, create a new one) under the opp_table_lock, so only
> > + * one creator gets access to the same. But that expands the critical section
> > + * under the lock and may end up causing circular dependencies with frameworks
> > + * like debugfs, interconnect or clock framework as they may be direct or
> > + * indirect users of OPP core.
> > + *
> > + * And for that reason we have to go for a bit tricky implementation here, which
> > + * uses the opp_tables_busy flag to indicate if another creator is in the middle
> > + * of adding an OPP table and others should wait for it to finish.
> > + */
> >  static struct opp_table *_opp_get_opp_table(struct device *dev, int index)
> >  {
> >       struct opp_table *opp_table;
> >
> > -     /* Hold our table modification lock here */
> > +again:
> >       mutex_lock(&opp_table_lock);
> >
> >       opp_table = _find_opp_table_unlocked(dev);
> >       if (!IS_ERR(opp_table))
> >               goto unlock;
> >
> > +     /*
> > +      * The opp_tables list or an OPP table's dev_list is getting updated by
> > +      * another user, wait for it to finish.
> > +      */
> > +     if (unlikely(opp_tables_busy)) {
> > +             mutex_unlock(&opp_table_lock);
> > +             cpu_relax();
> > +             goto again;
> > +     }
> > +
> > +     opp_tables_busy = true;
> >       opp_table = _managed_opp(dev, index);
> > +
> > +     /* Drop the lock to reduce the size of critical section */
> > +     mutex_unlock(&opp_table_lock);
> > +
> >       if (opp_table) {
> > -             if (!_add_opp_dev_unlocked(dev, opp_table)) {
> > +             if (!_add_opp_dev(dev, opp_table)) {
> >                       dev_pm_opp_put_opp_table(opp_table);
> >                       opp_table = ERR_PTR(-ENOMEM);
> >               }
> > -             goto unlock;
> > +
> > +             mutex_lock(&opp_table_lock);
> > +     } else {
> > +             opp_table = _allocate_opp_table(dev, index);
> > +
> > +             mutex_lock(&opp_table_lock);
> > +             if (!IS_ERR(opp_table))
> > +                     list_add(&opp_table->node, &opp_tables);
> >       }
> >
> > -     opp_table = _allocate_opp_table(dev, index);
> > +     opp_tables_busy = false;
> >
> >  unlock:
> >       mutex_unlock(&opp_table_lock);
> > @@ -1181,6 +1208,10 @@ static void _opp_table_kref_release(struct kref *kref)
> >       struct opp_device *opp_dev, *temp;
> >       int i;
> >
> > +     /* Drop the lock as soon as we can */
> > +     list_del(&opp_table->node);
> > +     mutex_unlock(&opp_table_lock);
> > +
> >       _of_clear_opp_table(opp_table);
> >
> >       /* Release clk */
> > @@ -1208,10 +1239,7 @@ static void _opp_table_kref_release(struct kref *kref)
> >
> >       mutex_destroy(&opp_table->genpd_virt_dev_lock);
> >       mutex_destroy(&opp_table->lock);
> > -     list_del(&opp_table->node);
> >       kfree(opp_table);
> > -
> > -     mutex_unlock(&opp_table_lock);
> >  }
> >
> >  void dev_pm_opp_put_opp_table(struct opp_table *opp_table)
>
> Rob, Ping.
>

sorry, it didn't apply cleanly (which I guess is due to some other
dependencies that need to be picked back to v5.4 product kernel), and
due to some other things I'm in middle of debugging I didn't have time
yet to switch to v5.10-rc or look at what else needs to
cherry-picked..

If you could, pushing a branch with this patch somewhere would be a
bit easier to work with (ie. fetch && cherry-pick is easier to deal
with than picking things from list)

BR,
-R