Le 08/09/2024 à 18:39, Biju Das a écrit :
Hi Christophe JAILLET,
Thanks for the patch.
-----Original Message-----
From: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
Sent: Saturday, September 7, 2024 2:59 PM
Subject: [PATCH v2] phy: renesas: rcar-gen3-usb2: Fix an error handling path in
rcar_gen3_phy_usb2_probe()
If an error occurs after the rcar_gen3_phy_usb2_init_bus(),
reset_control_assert() must be called, as already done in the remove function.
This is fine to re-use the existing error handling path, because even if "channel->rstc" is still NULL
at this point, it is safe to call reset_control_assert(NULL).
Fixes: 4eae16375357 ("phy: renesas: rcar-gen3-usb2: Add support to initialize the bus")
Signed-off-by: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
---
Changes in v2:
- Re-use 'error' to simplify the patch [claudiu beznea]
- Update the commit description to explain why it is safe.
v1:
https://lore.kernel.org/all/fc9f7b444f0ca645411868992bbe16514aeccfed.1725652654.git.christophe.jaillet
@wanadoo.fr/
---
drivers/phy/renesas/phy-rcar-gen3-usb2.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/phy/renesas/phy-rcar-gen3-usb2.c b/drivers/phy/renesas/phy-rcar-gen3-usb2.c
index 58e123305152..ccb0b54b70f7 100644
--- a/drivers/phy/renesas/phy-rcar-gen3-usb2.c
+++ b/drivers/phy/renesas/phy-rcar-gen3-usb2.c
@@ -803,6 +803,7 @@ static int rcar_gen3_phy_usb2_probe(struct platform_device *pdev)
return 0;
error:
+ reset_control_assert(channel->rstc);
This will result in either kernel crash [1] or reset usage count imbalance in case
of error [2] and [3] in rcar_gen3_phy_usb2_init_bus() see [4]. Also reset control API is not
respected for SoCs other than RZ/G3S. For those SoC's reset assert is
called without calling a get(). Maybe add a check (phy_data->init_bus) for
assert api's, that guarantees assert is called after calling a get() as it
valid only for RZ/G3S??
[1]
channel->rstc = devm_reset_control_array_get_shared(dev);
if (IS_ERR(channel->rstc))
return PTR_ERR(channel->rstc);
[2]
ret = pm_runtime_resume_and_get(dev);
if (ret)
return ret;
[3]
ret = reset_control_deassert(channel->rstc);
if (ret)
goto rpm_put;
[4] https://elixir.bootlin.com/linux/v6.11-rc6/source/drivers/reset/core.c#L483
So, if I understand correctly, v1 [5] was correct. :)
I don't think that [1] would crash, because of [6]. It would only
WARN_ON. But with v1, it is not called.
With v1, reset_control_assert() is not called if
rcar_gen3_phy_usb2_init_bus() fails. So [2] and [3] can't occur.
I can send a v3, which is the same of v1, or you can pick v1 as-is (if
I'm correct... :)) or you can just ignore it if "reset control API is
not respected for SoCs".
If of interest, I spotted it with one of my coccinelle script that
compares functions called in .remove function, but not in error handling
path of probe.
CJ
[5]:
https://lore.kernel.org/all/fc9f7b444f0ca645411868992bbe16514aeccfed.1725652654.git.christophe.jaillet@xxxxxxxxxx/
[6]:
https://elixir.bootlin.com/linux/v6.11-rc6/source/drivers/reset/core.c#L473
Cheers,
Biju
pm_runtime_disable(dev);
return ret;
--
2.46.0
