OSDN Git Service

net: sched: use rcu for action cookie update
authorVlad Buslov <vladbu@mellanox.com>
Thu, 5 Jul 2018 14:24:23 +0000 (17:24 +0300)
committerDavid S. Miller <davem@davemloft.net>
Sun, 8 Jul 2018 03:42:28 +0000 (12:42 +0900)
Implement functions to atomically update and free action cookie
using rcu mechanism.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/act_api.h
include/net/pkt_cls.h
net/sched/act_api.c

index 5ff11ad..ffc3ef3 100644 (file)
@@ -37,7 +37,7 @@ struct tc_action {
        spinlock_t                      tcfa_lock;
        struct gnet_stats_basic_cpu __percpu *cpu_bstats;
        struct gnet_stats_queue __percpu *cpu_qstats;
-       struct tc_cookie        *act_cookie;
+       struct tc_cookie        __rcu *act_cookie;
        struct tcf_chain        *goto_chain;
 };
 #define tcf_index      common.tcfa_index
index 6641584..2081e42 100644 (file)
@@ -781,6 +781,7 @@ struct tc_mqprio_qopt_offload {
 struct tc_cookie {
        u8  *data;
        u32 len;
+       struct rcu_head rcu;
 };
 
 struct tc_qopt_offload_stats {
index 3f4cf93..02670c7 100644 (file)
@@ -55,6 +55,24 @@ static void tcf_action_goto_chain_exec(const struct tc_action *a,
        res->goto_tp = rcu_dereference_bh(chain->filter_chain);
 }
 
+static void tcf_free_cookie_rcu(struct rcu_head *p)
+{
+       struct tc_cookie *cookie = container_of(p, struct tc_cookie, rcu);
+
+       kfree(cookie->data);
+       kfree(cookie);
+}
+
+static void tcf_set_action_cookie(struct tc_cookie __rcu **old_cookie,
+                                 struct tc_cookie *new_cookie)
+{
+       struct tc_cookie *old;
+
+       old = xchg(old_cookie, new_cookie);
+       if (old)
+               call_rcu(&old->rcu, tcf_free_cookie_rcu);
+}
+
 /* XXX: For standalone actions, we don't need a RCU grace period either, because
  * actions are always connected to filters and filters are already destroyed in
  * RCU callbacks, so after a RCU grace period actions are already disconnected
@@ -65,10 +83,7 @@ static void free_tcf(struct tc_action *p)
        free_percpu(p->cpu_bstats);
        free_percpu(p->cpu_qstats);
 
-       if (p->act_cookie) {
-               kfree(p->act_cookie->data);
-               kfree(p->act_cookie);
-       }
+       tcf_set_action_cookie(&p->act_cookie, NULL);
        if (p->goto_chain)
                tcf_action_goto_chain_fini(p);
 
@@ -567,16 +582,22 @@ tcf_action_dump_1(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
        int err = -EINVAL;
        unsigned char *b = skb_tail_pointer(skb);
        struct nlattr *nest;
+       struct tc_cookie *cookie;
 
        if (nla_put_string(skb, TCA_KIND, a->ops->kind))
                goto nla_put_failure;
        if (tcf_action_copy_stats(skb, a, 0))
                goto nla_put_failure;
-       if (a->act_cookie) {
-               if (nla_put(skb, TCA_ACT_COOKIE, a->act_cookie->len,
-                           a->act_cookie->data))
+
+       rcu_read_lock();
+       cookie = rcu_dereference(a->act_cookie);
+       if (cookie) {
+               if (nla_put(skb, TCA_ACT_COOKIE, cookie->len, cookie->data)) {
+                       rcu_read_unlock();
                        goto nla_put_failure;
+               }
        }
+       rcu_read_unlock();
 
        nest = nla_nest_start(skb, TCA_OPTIONS);
        if (nest == NULL)
@@ -719,13 +740,8 @@ struct tc_action *tcf_action_init_1(struct net *net, struct tcf_proto *tp,
        if (err < 0)
                goto err_mod;
 
-       if (name == NULL && tb[TCA_ACT_COOKIE]) {
-               if (a->act_cookie) {
-                       kfree(a->act_cookie->data);
-                       kfree(a->act_cookie);
-               }
-               a->act_cookie = cookie;
-       }
+       if (!name && tb[TCA_ACT_COOKIE])
+               tcf_set_action_cookie(&a->act_cookie, cookie);
 
        /* module count goes up only when brand new policy is created
         * if it exists and is only bound to in a_o->init() then