Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Alexandru Ardelean <aardelean@deviqon.com>
To:     linux-kernel@vger.kernel.org
Cc:     lgirdwood@gmail.com, broonie@kernel.org,
        Alexandru Ardelean <aardelean@deviqon.com>,
        Dmitry Torokhov <dmitry.torokhov@gmail.com>,
        Guenter Roeck <linux@roeck-us.net>
Subject: [RFC PATCH] regulator: devres: disable regulator on release if refcount is 1
Date:   Fri, 25 Jun 2021 15:53:07 +0300
Message-Id: <20210625125307.330831-1-aardelean@deviqon.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Evidently, this came about after doing a few of these types of constructs:

   static void reg_disable(void *reg)
   {
         regulator_disable(reg)
   }

   ...

   ret = regulator_enable(reg);
   if (ret)
       return ret;

   ret = devm_add_action_or_reset(dev, reg_disable, reg);

   ...

Naturally, the first thought was to try to move this construct into
regulator core, but I remembered that a devm_regulator_enable() function
isn't all that loved.
The construct above looks like it could become a short-hand (in the form of
devm_regulator_enable()), somewhere in the regulator framework.

After going back through the previous discussions [referenced below, sorry
if I missed any], it looks like maybe an idea would be to call
regulator_disable() right before regulator_put() inside
devm_regulator_release(). But we need to call it only if the 'enable_count'
is 1.

This means that the last 'regulator_disable()' (on driver remove) becomes
optional.
If there are any unbalanced regulator_enable()/regulator_disable() calls,
the 'enable_count' won't be touched and 'regulator_put()' will print a
warning.
The condition could be made to check if 'enable_count >= 1', and the
behavior would be the same, but it's probably a good idea not to touch this
refcount if isn't 1.

The only disadvantage to this approach, is that it changes the order in the
drivers in which the register_disable() gets called, with respect to other
steps in the probe/remove order.
With this, the register_disable() will be called right before the consumer
reference is free'd.

But the other advantage is that regulator_disable() calls can be removed
in simple drivers, where the consumer reference has been requested
with devm_regulator_get().

References:
  https://lore.kernel.org/lkml/20170213023249.GA27688@dtor-ws/
  https://lkml.org/lkml/2014/2/4/940

Cc: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Alexandru Ardelean <aardelean@deviqon.com>
---

Note: this patch applies indepently of this series:
  https://lore.kernel.org/lkml/20210625122324.327585-1-aardelean@deviqon.com/

 drivers/regulator/devres.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/drivers/regulator/devres.c b/drivers/regulator/devres.c
index 826c29499d69..1852afc02990 100644
--- a/drivers/regulator/devres.c
+++ b/drivers/regulator/devres.c
@@ -16,7 +16,12 @@
 
 static void devm_regulator_release(struct device *dev, void *res)
 {
-	regulator_put(*(struct regulator **)res);
+	struct regulator *regulator = *(struct regulator **)res;
+
+	if (regulator->enable_count == 1)
+		regulator_disable(regulator);
+
+	regulator_put(regulator);
 }
 
 static struct regulator *_devm_regulator_get(struct device *dev, const char *id,
-- 
2.31.1