tipc: eliminate message disordering during binding table update
authorJon Maloy <jon.maloy@ericsson.com>
Fri, 19 Oct 2018 17:55:40 +0000 (19:55 +0200)
committerDavid S. Miller <davem@davemloft.net>
Tue, 23 Oct 2018 02:29:12 +0000 (19:29 -0700)
We have seen the following race scenario:
1) named_distribute() builds a "bulk" message, containing a PUBLISH
   item for a certain publication. This is based on the contents of
   the binding tables's 'cluster_scope' list.
2) tipc_named_withdraw() removes the same publication from the list,
   bulds a WITHDRAW message and distributes it to all cluster nodes.
3) tipc_named_node_up(), which was calling named_distribute(), sends
   out the bulk message built under 1)
4) The WITHDRAW message arrives at the just detected node, finds
   no corresponding publication, and is dropped.
5) The PUBLISH item arrives at the same node, is added to its binding
   table, and remains there forever.

This arrival disordering was earlier taken care of by the backlog queue,
originally added for a different purpose, which was removed in the
commit referred to below, but we now need a different solution.
In this commit, we replace the rcu lock protecting the 'cluster_scope'
list with a regular RW lock which comprises even the sending of the
bulk message. This both guarantees both the list integrity and the
message sending order. We will later add a commit which cleans up
this code further.

Note that this commit needs recently added commit d3092b2efca1 ("tipc:
fix unsafe rcu locking when accessing publication list") to apply
cleanly.

Fixes: 37922ea4a310 ("tipc: permit overlapping service ranges in name table")
Reported-by: Tuong Lien Tong <tuong.t.lien@dektech.com.au>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/tipc/name_distr.c
net/tipc/name_table.c
net/tipc/name_table.h

index 3cfeb9df64b0008b894e7495cb0a00dd23938844..61219f0b9677fdd7934040a02d56248c43cf1a5e 100644 (file)
@@ -94,8 +94,9 @@ struct sk_buff *tipc_named_publish(struct net *net, struct publication *publ)
                list_add_tail_rcu(&publ->binding_node, &nt->node_scope);
                return NULL;
        }
                list_add_tail_rcu(&publ->binding_node, &nt->node_scope);
                return NULL;
        }
-       list_add_tail_rcu(&publ->binding_node, &nt->cluster_scope);
-
+       write_lock_bh(&nt->cluster_scope_lock);
+       list_add_tail(&publ->binding_node, &nt->cluster_scope);
+       write_unlock_bh(&nt->cluster_scope_lock);
        skb = named_prepare_buf(net, PUBLICATION, ITEM_SIZE, 0);
        if (!skb) {
                pr_warn("Publication distribution failure\n");
        skb = named_prepare_buf(net, PUBLICATION, ITEM_SIZE, 0);
        if (!skb) {
                pr_warn("Publication distribution failure\n");
@@ -112,11 +113,13 @@ struct sk_buff *tipc_named_publish(struct net *net, struct publication *publ)
  */
 struct sk_buff *tipc_named_withdraw(struct net *net, struct publication *publ)
 {
  */
 struct sk_buff *tipc_named_withdraw(struct net *net, struct publication *publ)
 {
+       struct name_table *nt = tipc_name_table(net);
        struct sk_buff *buf;
        struct distr_item *item;
 
        struct sk_buff *buf;
        struct distr_item *item;
 
-       list_del_rcu(&publ->binding_node);
-
+       write_lock_bh(&nt->cluster_scope_lock);
+       list_del(&publ->binding_node);
+       write_unlock_bh(&nt->cluster_scope_lock);
        if (publ->scope == TIPC_NODE_SCOPE)
                return NULL;
 
        if (publ->scope == TIPC_NODE_SCOPE)
                return NULL;
 
@@ -147,7 +150,7 @@ static void named_distribute(struct net *net, struct sk_buff_head *list,
                        ITEM_SIZE) * ITEM_SIZE;
        u32 msg_rem = msg_dsz;
 
                        ITEM_SIZE) * ITEM_SIZE;
        u32 msg_rem = msg_dsz;
 
-       list_for_each_entry_rcu(publ, pls, binding_node) {
+       list_for_each_entry(publ, pls, binding_node) {
                /* Prepare next buffer: */
                if (!skb) {
                        skb = named_prepare_buf(net, PUBLICATION, msg_rem,
                /* Prepare next buffer: */
                if (!skb) {
                        skb = named_prepare_buf(net, PUBLICATION, msg_rem,
@@ -189,11 +192,10 @@ void tipc_named_node_up(struct net *net, u32 dnode)
 
        __skb_queue_head_init(&head);
 
 
        __skb_queue_head_init(&head);
 
-       rcu_read_lock();
+       read_lock_bh(&nt->cluster_scope_lock);
        named_distribute(net, &head, dnode, &nt->cluster_scope);
        named_distribute(net, &head, dnode, &nt->cluster_scope);
-       rcu_read_unlock();
-
        tipc_node_xmit(net, &head, dnode, 0);
        tipc_node_xmit(net, &head, dnode, 0);
+       read_unlock_bh(&nt->cluster_scope_lock);
 }
 
 /**
 }
 
 /**
index 66d5b2c5987ad6a2e05ca45f236ade58459ac0bf..bff241f0352501aba8605622df16f2c85044c09b 100644 (file)
@@ -744,6 +744,7 @@ int tipc_nametbl_init(struct net *net)
 
        INIT_LIST_HEAD(&nt->node_scope);
        INIT_LIST_HEAD(&nt->cluster_scope);
 
        INIT_LIST_HEAD(&nt->node_scope);
        INIT_LIST_HEAD(&nt->cluster_scope);
+       rwlock_init(&nt->cluster_scope_lock);
        tn->nametbl = nt;
        spin_lock_init(&tn->nametbl_lock);
        return 0;
        tn->nametbl = nt;
        spin_lock_init(&tn->nametbl_lock);
        return 0;
index 892bd750b85fa48639b15af9bbeb6d437a7660a7..f79066334cc8f63a4bf71a61955b155c66b019be 100644 (file)
@@ -100,6 +100,7 @@ struct name_table {
        struct hlist_head services[TIPC_NAMETBL_SIZE];
        struct list_head node_scope;
        struct list_head cluster_scope;
        struct hlist_head services[TIPC_NAMETBL_SIZE];
        struct list_head node_scope;
        struct list_head cluster_scope;
+       rwlock_t cluster_scope_lock;
        u32 local_publ_count;
 };
 
        u32 local_publ_count;
 };