It has turned out that having _set_required_opps() to recursively call
dev_pm_opp_set_opp() to set the required OPPs, doesn't really work as well
as we expected.
More precisely, at each recursive call to dev_pm_opp_set_opp() we are
changing an OPP for a required_dev that belongs to a required-OPP table.
The problem with this, is that we may have several devices sharing the same
required-OPP table, which leads to an incorrect behaviour in regards to
aggregating the per device votes.
To fix the problem for a required-OPP table belonging to a PM domain, which
is the only existing usecase for now, let's simply replace the call to
dev_pm_opp_set_opp() in _set_required_opps() by a call to _set_opp_level().
Moving forward we may potentially need to add support for other types of
required-OPP tables. In this case, the aggregation needs to be thought of.
Fixes: e37440e7e2c2 ("OPP: Call dev_pm_opp_set_opp() for required OPPs")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
Acked-by: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
Link: https://lore.kernel.org/r/20240822224547.385095-2-ulf.hansson@xxxxxxxxxx
diff --git a/drivers/opp/core.c b/drivers/opp/core.c
index 5f4598246a87..494f8860220d 100644
--- a/drivers/opp/core.c
+++ b/drivers/opp/core.c
@@ -1061,6 +1061,27 @@ static int _set_opp_bw(const struct opp_table *opp_table,
return 0;
}
+static int _set_opp_level(struct device *dev, struct dev_pm_opp *opp)
+{
+ unsigned int level = 0;
+ int ret = 0;
+
+ if (opp) {
+ if (opp->level == OPP_LEVEL_UNSET)
+ return 0;
+
+ level = opp->level;
+ }
+
+ /* Request a new performance state through the device's PM domain. */
+ ret = dev_pm_domain_set_performance_state(dev, level);
+ if (ret)
+ dev_err(dev, "Failed to set performance state %u (%d)\n", level,
+ ret);
+
+ return ret;
+}
+
/* This is only called for PM domain for now */
static int _set_required_opps(struct device *dev, struct opp_table *opp_table,
struct dev_pm_opp *opp, bool up)
@@ -1091,7 +1112,7 @@ static int _set_required_opps(struct device *dev, struct opp_table *opp_table,
if (devs[index]) {
required_opp = opp ? opp->required_opps[index] : NULL;
- ret = dev_pm_opp_set_opp(devs[index], required_opp);
+ ret = _set_opp_level(devs[index], required_opp);
if (ret)
return ret;
}
@@ -1102,27 +1123,6 @@ static int _set_required_opps(struct device *dev, struct opp_table *opp_table,
return 0;
}
-static int _set_opp_level(struct device *dev, struct dev_pm_opp *opp)
-{
- unsigned int level = 0;
- int ret = 0;
-
- if (opp) {
- if (opp->level == OPP_LEVEL_UNSET)
- return 0;
-
- level = opp->level;
- }
-
- /* Request a new performance state through the device's PM domain. */
- ret = dev_pm_domain_set_performance_state(dev, level);
- if (ret)
- dev_err(dev, "Failed to set performance state %u (%d)\n", level,
- ret);
-
- return ret;
-}
-
static void _find_current_opp(struct device *dev, struct opp_table *opp_table)
{
struct dev_pm_opp *opp = ERR_PTR(-ENODEV);
@@ -2457,18 +2457,6 @@ static int _opp_attach_genpd(struct opp_table *opp_table, struct device *dev,
}
}
- /*
- * Add the virtual genpd device as a user of the OPP table, so
- * we can call dev_pm_opp_set_opp() on it directly.
- *
- * This will be automatically removed when the OPP table is
- * removed, don't need to handle that here.
- */
- if (!_add_opp_dev(virt_dev, opp_table->required_opp_tables[index])) {
- ret = -ENOMEM;
- goto err;
- }
-
opp_table->required_devs[index] = virt_dev;
index++;
name++;
