Thread (15 messages) flat view 15 messages, 3 authors, 2021-01-26

Re: [PATCH v6 net-next 07/10] net: dsa: allow changing the tag protocol via the "tagging" device attribute

From: Vladimir Oltean <olteanv@gmail.com>
Date: 2021-01-24 08:21:47

On Fri, Jan 22, 2021 at 08:52:16PM -0800, Jakub Kicinski wrote:
On Thu, 21 Jan 2021 18:01:28 +0200 Vladimir Oltean wrote:
quoted
+	/* At the moment we don't allow changing the tag protocol under
+	 * traffic. May revisit in the future.
+	 */
+	if (master->flags & IFF_UP)
+		return -EBUSY;
But you're not holding rtnl_lock at this point, this check is advisory
at best.
Yes, I should hold the rtnl_mutex.
quoted
+	list_for_each_entry(dp, &dst->ports, list) {
What protects this iteration? All sysfs guarantees you is that
struct net_device *master itself will not disappear.

Could you explain the locking expectations a bit?
The dsa_group sysfs is removed in:

dsa_unregister_switch
-> mutex_lock(&dsa2_mutex)
-> dsa_switch_remove
   -> dsa_tree_teardown
      -> dsa_tree_teardown_master
         -> dsa_master_teardown
            -> sysfs_remove_group
There are 2 points here:
1. sysfs_remove_group actually waits for a concurrent tagging_store()
   call to finish (at least it does when I put an msleep(10000) inside
   tagging_store).
2. After the sysfs_remove_group, dsa_tree_change_tag_proto should never
   be called again.

Next comes:
   -> dsa_tree_teardown
      -> dsa_tree_teardown_switches
         -> dsa_port_teardown
            -> dsa_slave_destroy
After this, all DSA net devices are unregistered and freed.

Next comes:
-> dsa_switch_remove
   -> dsa_switch_release_ports
-> mutex_unlock(&dsa2_mutex)
where the dst->ports list is finally freed.

So there is no chance that the dst->ports list is modified concurrently
with tagging_store.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help