net: sched: cls_cgroup use RCU
authorJohn Fastabend <john.fastabend@gmail.com>
Sat, 13 Sep 2014 03:06:26 +0000 (20:06 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sat, 13 Sep 2014 16:29:59 +0000 (12:29 -0400)
Make cgroup classifier safe for RCU.

Also drops the calls in the classify routine that were doing a
rcu_read_lock()/rcu_read_unlock(). If the rcu_read_lock() isn't held
entering this routine we have issues with deleting the classifier
chain so remove the unnecessary rcu_read_lock()/rcu_read_unlock()
pair noting all paths AFAIK hold rcu_read_lock.

If there is a case where classify is called without the rcu read lock
then an rcu splat will occur and we can correct it.

Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/sched/cls_cgroup.c

index cacf01bd04f0a96660050c0e71da0840c8af0f95..3b7548759998b10e0f6262dda7c910611ec9f89f 100644 (file)
@@ -22,17 +22,17 @@ struct cls_cgroup_head {
        u32                     handle;
        struct tcf_exts         exts;
        struct tcf_ematch_tree  ematches;
+       struct tcf_proto        *tp;
+       struct rcu_head         rcu;
 };
 
 static int cls_cgroup_classify(struct sk_buff *skb, const struct tcf_proto *tp,
                               struct tcf_result *res)
 {
-       struct cls_cgroup_head *head = tp->root;
+       struct cls_cgroup_head *head = rcu_dereference_bh(tp->root);
        u32 classid;
 
-       rcu_read_lock();
        classid = task_cls_state(current)->classid;
-       rcu_read_unlock();
 
        /*
         * Due to the nature of the classifier it is required to ignore all
@@ -80,13 +80,25 @@ static const struct nla_policy cgroup_policy[TCA_CGROUP_MAX + 1] = {
        [TCA_CGROUP_EMATCHES]   = { .type = NLA_NESTED },
 };
 
+static void cls_cgroup_destroy_rcu(struct rcu_head *root)
+{
+       struct cls_cgroup_head *head = container_of(root,
+                                                   struct cls_cgroup_head,
+                                                   rcu);
+
+       tcf_exts_destroy(head->tp, &head->exts);
+       tcf_em_tree_destroy(head->tp, &head->ematches);
+       kfree(head);
+}
+
 static int cls_cgroup_change(struct net *net, struct sk_buff *in_skb,
                             struct tcf_proto *tp, unsigned long base,
                             u32 handle, struct nlattr **tca,
                             unsigned long *arg, bool ovr)
 {
        struct nlattr *tb[TCA_CGROUP_MAX + 1];
-       struct cls_cgroup_head *head = tp->root;
+       struct cls_cgroup_head *head = rtnl_dereference(tp->root);
+       struct cls_cgroup_head *new;
        struct tcf_ematch_tree t;
        struct tcf_exts e;
        int err;
@@ -94,25 +106,24 @@ static int cls_cgroup_change(struct net *net, struct sk_buff *in_skb,
        if (!tca[TCA_OPTIONS])
                return -EINVAL;
 
-       if (head == NULL) {
-               if (!handle)
-                       return -EINVAL;
+       if (!head && !handle)
+               return -EINVAL;
 
-               head = kzalloc(sizeof(*head), GFP_KERNEL);
-               if (head == NULL)
-                       return -ENOBUFS;
+       if (head && handle != head->handle)
+               return -ENOENT;
 
-               tcf_exts_init(&head->exts, TCA_CGROUP_ACT, TCA_CGROUP_POLICE);
-               head->handle = handle;
+       new = kzalloc(sizeof(*head), GFP_KERNEL);
+       if (!new)
+               return -ENOBUFS;
 
-               tcf_tree_lock(tp);
-               tp->root = head;
-               tcf_tree_unlock(tp);
+       if (head) {
+               new->handle = head->handle;
+       } else {
+               tcf_exts_init(&new->exts, TCA_CGROUP_ACT, TCA_CGROUP_POLICE);
+               new->handle = handle;
        }
 
-       if (handle != head->handle)
-               return -ENOENT;
-
+       new->tp = tp;
        err = nla_parse_nested(tb, TCA_CGROUP_MAX, tca[TCA_OPTIONS],
                               cgroup_policy);
        if (err < 0)
@@ -127,20 +138,24 @@ static int cls_cgroup_change(struct net *net, struct sk_buff *in_skb,
        if (err < 0)
                return err;
 
-       tcf_exts_change(tp, &head->exts, &e);
-       tcf_em_tree_change(tp, &head->ematches, &t);
+       tcf_exts_change(tp, &new->exts, &e);
+       tcf_em_tree_change(tp, &new->ematches, &t);
 
+       rcu_assign_pointer(tp->root, new);
+       if (head)
+               call_rcu(&head->rcu, cls_cgroup_destroy_rcu);
        return 0;
 }
 
 static void cls_cgroup_destroy(struct tcf_proto *tp)
 {
-       struct cls_cgroup_head *head = tp->root;
+       struct cls_cgroup_head *head = rtnl_dereference(tp->root);
 
        if (head) {
                tcf_exts_destroy(tp, &head->exts);
                tcf_em_tree_destroy(tp, &head->ematches);
-               kfree(head);
+               RCU_INIT_POINTER(tp->root, NULL);
+               kfree_rcu(head, rcu);
        }
 }
 
@@ -151,7 +166,7 @@ static int cls_cgroup_delete(struct tcf_proto *tp, unsigned long arg)
 
 static void cls_cgroup_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 {
-       struct cls_cgroup_head *head = tp->root;
+       struct cls_cgroup_head *head = rtnl_dereference(tp->root);
 
        if (arg->count < arg->skip)
                goto skip;
@@ -167,7 +182,7 @@ skip:
 static int cls_cgroup_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
                           struct sk_buff *skb, struct tcmsg *t)
 {
-       struct cls_cgroup_head *head = tp->root;
+       struct cls_cgroup_head *head = rtnl_dereference(tp->root);
        unsigned char *b = skb_tail_pointer(skb);
        struct nlattr *nest;