[B.A.T.M.A.N.] [PATCH] batman-adv: remove obsolete deleted attribute for gateway node

[B.A.T.M.A.N.] [PATCH] batman-adv: remove obsolete deleted attribute for gateway node

[Simon Wunderlich]

From: Simon Wunderlich <simon@open-mesh.com>

With rcu, the gateway node deleted attribute is not needed anymore. In
fact, it may delay the free of the gateway node and its referenced
structures. Therefore remove it altogether and simplify purging as well.

Signed-off-by: Simon Wunderlich <simon@open-mesh.com>
---
 net/batman-adv/gateway_client.c | 39 +++++++++------------------------------
 net/batman-adv/originator.c     |  1 -
 net/batman-adv/types.h          |  2 --
 3 files changed, 9 insertions(+), 33 deletions(-)

diff --git a/net/batman-adv/gateway_client.c b/net/batman-adv/gateway_client.c
index 97f7c44..ca436ff 100644
--- a/net/batman-adv/gateway_client.c
+++ b/net/batman-adv/gateway_client.c
@@ -165,9 +165,6 @@ batadv_gw_get_best_gw_node(struct batadv_priv *bat_priv)
 
 	rcu_read_lock();
 	hlist_for_each_entry_rcu(gw_node, &bat_priv->gw.list, list) {
-		if (gw_node->deleted)
-			continue;
-
 		orig_node = gw_node->orig_node;
 		router = batadv_orig_router_get(orig_node, BATADV_IF_DEFAULT);
 		if (!router)
@@ -475,9 +472,6 @@ batadv_gw_node_get(struct batadv_priv *bat_priv,
 		if (gw_node_tmp->orig_node != orig_node)
 			continue;
 
-		if (gw_node_tmp->deleted)
-			continue;
-
 		if (!atomic_inc_not_zero(&gw_node_tmp->refcount))
 			continue;
 
@@ -527,9 +521,7 @@ void batadv_gw_node_update(struct batadv_priv *bat_priv,
 	gw_node->bandwidth_down = ntohl(gateway->bandwidth_down);
 	gw_node->bandwidth_up = ntohl(gateway->bandwidth_up);
 
-	gw_node->deleted = 0;
 	if (ntohl(gateway->bandwidth_down) == 0) {
-		gw_node->deleted = jiffies;
 		batadv_dbg(BATADV_DBG_BATMAN, bat_priv,
 			   "Gateway %pM removed from gateway list\n",
 			   orig_node->orig);
@@ -537,9 +529,15 @@ void batadv_gw_node_update(struct batadv_priv *bat_priv,
 		/* Note: We don't need a NULL check here, since curr_gw never
 		 * gets dereferenced.
 		 */
+		spin_lock_bh(&bat_priv->gw.list_lock);
+		hlist_del_rcu(&gw_node->list);
+		spin_unlock_bh(&bat_priv->gw.list_lock);
+
 		curr_gw = batadv_gw_get_selected_gw_node(bat_priv);
 		if (gw_node == curr_gw)
 			batadv_gw_reselect(bat_priv);
+
+		batadv_gw_node_free_ref(gw_node);
 	}
 
 out:
@@ -562,37 +560,21 @@ void batadv_gw_node_delete(struct batadv_priv *bat_priv,
 
 void batadv_gw_node_purge(struct batadv_priv *bat_priv)
 {
-	struct batadv_gw_node *gw_node, *curr_gw;
+	struct batadv_gw_node *gw_node;
 	struct hlist_node *node_tmp;
-	unsigned long timeout = msecs_to_jiffies(2 * BATADV_PURGE_TIMEOUT);
-	int do_reselect = 0;
 
-	curr_gw = batadv_gw_get_selected_gw_node(bat_priv);
+	if (atomic_read(&bat_priv->mesh_state) == BATADV_MESH_ACTIVE)
+		return;
 
 	spin_lock_bh(&bat_priv->gw.list_lock);
-
 	hlist_for_each_entry_safe(gw_node, node_tmp,
 				  &bat_priv->gw.list, list) {
-		if (((!gw_node->deleted) ||
-		     (time_before(jiffies, gw_node->deleted + timeout))) &&
-		    atomic_read(&bat_priv->mesh_state) == BATADV_MESH_ACTIVE)
-			continue;
-
-		if (curr_gw == gw_node)
-			do_reselect = 1;
 
 		hlist_del_rcu(&gw_node->list);
 		batadv_gw_node_free_ref(gw_node);
 	}
 
 	spin_unlock_bh(&bat_priv->gw.list_lock);
-
-	/* gw_reselect() needs to acquire the gw_list_lock */
-	if (do_reselect)
-		batadv_gw_reselect(bat_priv);
-
-	if (curr_gw)
-		batadv_gw_node_free_ref(curr_gw);
 }
 
 /* fails if orig_node has no router */
@@ -656,9 +638,6 @@ int batadv_gw_client_seq_print_text(struct seq_file *seq, void *offset)
 
 	rcu_read_lock();
 	hlist_for_each_entry_rcu(gw_node, &bat_priv->gw.list, list) {
-		if (gw_node->deleted)
-			continue;
-
 		/* fails if orig_node has no router */
 		if (batadv_write_buffer_text(bat_priv, seq, gw_node) < 0)
 			continue;
diff --git a/net/batman-adv/originator.c b/net/batman-adv/originator.c
index 4500e3a..016fc20 100644
--- a/net/batman-adv/originator.c
+++ b/net/batman-adv/originator.c
@@ -1009,7 +1009,6 @@ static void _batadv_purge_orig(struct batadv_priv *bat_priv)
 		spin_unlock_bh(list_lock);
 	}
 
-	batadv_gw_node_purge(bat_priv);
 	batadv_gw_election(bat_priv);
 }
 
diff --git a/net/batman-adv/types.h b/net/batman-adv/types.h
index da4c738..002bca4 100644
--- a/net/batman-adv/types.h
+++ b/net/batman-adv/types.h
@@ -328,7 +328,6 @@ enum batadv_orig_capabilities {
  * @orig_node: pointer to corresponding orig node
  * @bandwidth_down: advertised uplink download bandwidth
  * @bandwidth_up: advertised uplink upload bandwidth
- * @deleted: this struct is scheduled for deletion
  * @refcount: number of contexts the object is used
  * @rcu: struct used for freeing in an RCU-safe manner
  */
@@ -337,7 +336,6 @@ struct batadv_gw_node {
 	struct batadv_orig_node *orig_node;
 	u32 bandwidth_down;
 	u32 bandwidth_up;
-	unsigned long deleted;
 	atomic_t refcount;
 	struct rcu_head rcu;
 };
-- 
2.1.4

Re: [B.A.T.M.A.N.] [PATCH] batman-adv: remove obsolete deleted attribute for gateway node

[Antonio Quartulli]

[-- Attachment #1: Type: text/plain, Size: 710 bytes --]



On 24/06/15 14:50, Simon Wunderlich wrote:
> @@ -537,9 +529,15 @@ void batadv_gw_node_update(struct batadv_priv *bat_priv,
>  		/* Note: We don't need a NULL check here, since curr_gw never
>  		 * gets dereferenced.
>  		 */
> +		spin_lock_bh(&bat_priv->gw.list_lock);
> +		hlist_del_rcu(&gw_node->list);
> +		spin_unlock_bh(&bat_priv->gw.list_lock);
> +

have you verified that we can't get to this point through two threads
and enter the same problem that Sven and Marek have been discussing and
fixing in these days: removing an element from a list twice

If this is possible, you should check that this element is in the list
before deleting it.

Cheers,

-- 
Antonio Quartulli


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: [B.A.T.M.A.N.] [PATCH] batman-adv: remove obsolete deleted attribute for gateway node

[Simon Wunderlich]

On Wednesday 24 June 2015 19:00:18 Antonio Quartulli wrote:
> On 24/06/15 14:50, Simon Wunderlich wrote:
> > @@ -537,9 +529,15 @@ void batadv_gw_node_update(struct batadv_priv
> > *bat_priv,> 
> >  		/* Note: We don't need a NULL check here, since curr_gw never
> >  		
> >  		 * gets dereferenced.
> >  		 */
> > 
> > +		spin_lock_bh(&bat_priv->gw.list_lock);
> > +		hlist_del_rcu(&gw_node->list);
> > +		spin_unlock_bh(&bat_priv->gw.list_lock);
> > +
> 
> have you verified that we can't get to this point through two threads
> and enter the same problem that Sven and Marek have been discussing and
> fixing in these days: removing an element from a list twice
> 
> If this is possible, you should check that this element is in the list
> before deleting it.

That's an excellent point, I didn't check that. It also appears a little 
complicated, so I'll add some more protection at this point and resend v2.

Thank you!
     Simon