[PATCH] [RFC] soc: imx: gpcv2: Split clock prepare from clock enable in the domain

[Marek Vasut]

It is possible for clk_disable_unused() to trigger lockdep warning
regarding lock ordering in this driver. This happens in case of the
following conditions:

A) clock core clk_disable_unused() triggers the following sequence in a
   driver which also uses GPCv2 domain:
   - clk_prepare_lock() -> obtains clock core prepare_lock
   - pm_runtime_get*() -> obtains &blk_ctrl_genpd_lock_class

B) driver powers up a power domain and triggers the following sequence
   in GPCv2:
   - pm_runtime_get_sync() -> obtains &blk_ctrl_genpd_lock_class
   - clk_bulk_prepare_enable() -> obtains clock core prepare_lock

This can lead to a deadlock in case A and B runs on separate CPUs.

To avoid the deadlock, split clk_*prepare() from clk_*enable() and
call the former before pm_runtime_get_sync(). This way, the GPCv2
driver always claims the prepare_lock before blk_ctrl_genpd_lock_class
and the deadlock is avoided.

The imx8m and imx8mp block controller drivers likely need similar fix.

Signed-off-by: Marek Vasut <marex at denx.de>
---
Cc: Fabio Estevam <festevam at denx.de>
Cc: Jacky Bai <ping.bai at nxp.com>
Cc: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
Cc: Lucas Stach <l.stach at pengutronix.de>
Cc: Peng Fan <peng.fan at nxp.com>
Cc: Shawn Guo <shawnguo at kernel.org>
Cc: Shengjiu Wang <shengjiu.wang at nxp.com>
Cc: Stephen Boyd <sboyd at kernel.org>
Cc: linux-clk at vger.kernel.org
Cc: linux-imx at nxp.com
To: linux-arm-kernel at lists.infradead.org
---
 drivers/soc/imx/gpcv2.c | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/drivers/soc/imx/gpcv2.c b/drivers/soc/imx/gpcv2.c
index 88aee59730e39..ee9294ad25ba1 100644
--- a/drivers/soc/imx/gpcv2.c
+++ b/drivers/soc/imx/gpcv2.c
@@ -319,10 +319,16 @@ static int imx_pgc_power_up(struct generic_pm_domain *genpd)
 	u32 reg_val, pgc;
 	int ret;
 
+	ret = clk_bulk_prepare(domain->num_clks, domain->clks);
+	if (ret) {
+		dev_err(domain->dev, "failed to prepare reset clocks\n");
+		return ret;
+	}
+
 	ret = pm_runtime_get_sync(domain->dev);
 	if (ret < 0) {
 		pm_runtime_put_noidle(domain->dev);
-		return ret;
+		goto out_clock_unprepare;
 	}
 
 	if (!IS_ERR(domain->regulator)) {
@@ -338,7 +344,7 @@ static int imx_pgc_power_up(struct generic_pm_domain *genpd)
 	reset_control_assert(domain->reset);
 
 	/* Enable reset clocks for all devices in the domain */
-	ret = clk_bulk_prepare_enable(domain->num_clks, domain->clks);
+	ret = clk_bulk_enable(domain->num_clks, domain->clks);
 	if (ret) {
 		dev_err(domain->dev, "failed to enable reset clocks\n");
 		goto out_regulator_disable;
@@ -402,12 +408,14 @@ static int imx_pgc_power_up(struct generic_pm_domain *genpd)
 	return 0;
 
 out_clk_disable:
-	clk_bulk_disable_unprepare(domain->num_clks, domain->clks);
+	clk_bulk_disable(domain->num_clks, domain->clks);
 out_regulator_disable:
 	if (!IS_ERR(domain->regulator))
 		regulator_disable(domain->regulator);
 out_put_pm:
 	pm_runtime_put(domain->dev);
+out_clock_unprepare:
+	clk_bulk_unprepare(domain->num_clks, domain->clks);
 
 	return ret;
 }
-- 
2.35.1