On tor, mar 10, 2022 at 17:07, Vladimir Oltean [off-list ref] wrote:

On Thu, Mar 10, 2022 at 04:00:52PM +0100, Hans Schultz wrote:

quoted

quoted

quoted

+	brport = dsa_port_to_bridge_port(dp);

Since this is threaded interrupt context, I suppose it could race with
dsa_port_bridge_leave(). So it is best to check whether "brport" is NULL
or not.

Would something like:
if (dsa_is_unused_port(chip->ds, port))
        return -ENODATA;

be appropriate and sufficient for that?

static inline
struct net_device *dsa_port_to_bridge_port(const struct dsa_port *dp)
{
	if (!dp->bridge)
		return NULL;

	if (dp->lag)
		return dp->lag->dev;
	else if (dp->hsr_dev)
		return dp->hsr_dev;

	return dp->slave;
}

Notice the "dp->bridge" check. The assignments are in dsa_port_bridge_create()
and in dsa_port_bridge_destroy(). These functions assume rtnl_mutex protection.
The question was how do you serialize with that, and why do you assume
that dsa_port_to_bridge_port() returns non-NULL.

So no, dsa_is_unused_port() would do absolutely nothing to help.

I was thinking in indirect terms (dangerous I know :-).

But wrt the nl lock, I wonder when other threads could pull the carpet
away under this, and so I might have to wait till after the last call
(mv88e6xxx_g1_atu_loadpurge) to free the nl lock?