Re: [PATCH net 2/4] net: ethernet: ti: cpsw: fix device and of_node leaks

On Tue, Nov 01, 2016 at 12:48:48PM -0400, David Miller wrote:

From: Johan Hovold <johan@kernel.org>
Date: Tue, 1 Nov 2016 17:42:25 +0100

quoted

On Tue, Nov 01, 2016 at 12:27:11PM -0400, David Miller wrote:

quoted

From: Johan Hovold <johan@kernel.org>
Date: Tue,  1 Nov 2016 12:03:35 +0100

quoted

diff --git a/drivers/net/ethernet/ti/cpsw-phy-sel.c b/drivers/net/ethernet/ti/cpsw-phy-sel.c
index 054a8dd23dae..589beb843f56 100644
--- a/drivers/net/ethernet/ti/cpsw-phy-sel.c
+++ b/drivers/net/ethernet/ti/cpsw-phy-sel.c

@@ -176,8 +176,11 @@ void cpsw_phy_sel(struct device *dev, phy_interface_t phy_mode, int slave)
 	}
 
 	dev = bus_find_device(&platform_bus_type, NULL, node, match);
+	of_node_put(node);
 	priv = dev_get_drvdata(dev);
 
+	put_device(dev);
+
 	priv->cpsw_phy_sel(priv, phy_mode, slave);
 }
 EXPORT_SYMBOL_GPL(cpsw_phy_sel);

The only reference you have to 'dev' is the one obtained from the
bus_find_device() call, therefore you must at least hold onto
'dev' until after the priv->cpsw_phy_sel(priv, phy_mode, slave); call.

As I mentioned in the commit message "...there is no guarantee that the
devres-managed struct cpsw_phy_sel_priv will continue to be valid until
this function returns regardless of this change".

Specifically, holding a reference to dev does not prevent the
cpsw_phy_sel driver from being unbound and priv from being freed.

But you should at least hold onto the object while you call a function
pointer embedded in a data structure referred by it.

But there is no such reference to be held (currently). While priv is
valid (i.e. dev has a driver bound), driver core will hold a reference
to dev. If someone unbinds the driver, priv will be gone long before dev
and all bets are off anyway.

So I released the reference to dev before dereferencing priv on purpose
to avoid having someone make false assumptions about the lifetime of priv.

Johan