net_sched: act_bpf: remove spinlock in fast path
authorAlexei Starovoitov <ast@plumgrid.com>
Wed, 26 Aug 2015 03:06:35 +0000 (20:06 -0700)
committerDavid S. Miller <davem@davemloft.net>
Wed, 26 Aug 2015 18:01:45 +0000 (11:01 -0700)
Similar to act_gact/act_mirred, act_bpf can be lockless in packet processing
with extra care taken to free bpf programs after rcu grace period.
Replacement of existing act_bpf (very rare) is done with synchronize_rcu()
and final destruction is done from tc_action_ops->cleanup() callback that is
called from tcf_exts_destroy()->tcf_action_destroy()->__tcf_hash_release() when
bind and refcnt reach zero which is only possible when classifier is destroyed.
Previous two patches fixed the last two classifiers (tcindex and rsvp) to
call tcf_exts_destroy() from rcu callback.

Similar to gact/mirred there is a race between prog->filter and
prog->tcf_action. Meaning that the program being replaced may use
previous default action if it happened to return TC_ACT_UNSPEC.
act_mirred race betwen tcf_action and tcfm_dev is similar.
In all cases the race is harmless.
Long term we may want to improve the situation by replacing the whole
tc_action->priv as single pointer instead of updating inner fields one by one.

Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tc_act/tc_bpf.h
net/sched/act_bpf.c

index a152e9858b2c819a8429cb0ebd5a2af11690e007..958d69cfb19cc1fa5fed2f3fd64b5a2432f978b6 100644 (file)
@@ -15,7 +15,7 @@
 
 struct tcf_bpf {
        struct tcf_common       common;
-       struct bpf_prog         *filter;
+       struct bpf_prog __rcu   *filter;
        union {
                u32             bpf_fd;
                u16             bpf_num_ops;
index 458cf647e698ae46ec8f394fa6c94e0d77345d68..559bfa011bda7b7c6fed76ef7df8d4cd1d861bc3 100644 (file)
@@ -37,25 +37,24 @@ static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act,
                   struct tcf_result *res)
 {
        struct tcf_bpf *prog = act->priv;
+       struct bpf_prog *filter;
        int action, filter_res;
        bool at_ingress = G_TC_AT(skb->tc_verd) & AT_INGRESS;
 
        if (unlikely(!skb_mac_header_was_set(skb)))
                return TC_ACT_UNSPEC;
 
-       spin_lock(&prog->tcf_lock);
-
-       prog->tcf_tm.lastuse = jiffies;
-       bstats_update(&prog->tcf_bstats, skb);
+       tcf_lastuse_update(&prog->tcf_tm);
+       bstats_cpu_update(this_cpu_ptr(prog->common.cpu_bstats), skb);
 
-       /* Needed here for accessing maps. */
        rcu_read_lock();
+       filter = rcu_dereference(prog->filter);
        if (at_ingress) {
                __skb_push(skb, skb->mac_len);
-               filter_res = BPF_PROG_RUN(prog->filter, skb);
+               filter_res = BPF_PROG_RUN(filter, skb);
                __skb_pull(skb, skb->mac_len);
        } else {
-               filter_res = BPF_PROG_RUN(prog->filter, skb);
+               filter_res = BPF_PROG_RUN(filter, skb);
        }
        rcu_read_unlock();
 
@@ -77,7 +76,7 @@ static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act,
                break;
        case TC_ACT_SHOT:
                action = filter_res;
-               prog->tcf_qstats.drops++;
+               qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats));
                break;
        case TC_ACT_UNSPEC:
                action = prog->tcf_action;
@@ -87,7 +86,6 @@ static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act,
                break;
        }
 
-       spin_unlock(&prog->tcf_lock);
        return action;
 }
 
@@ -263,7 +261,10 @@ static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
                                  struct tcf_bpf_cfg *cfg)
 {
        cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
-       cfg->filter = prog->filter;
+       /* updates to prog->filter are prevented, since it's called either
+        * with rtnl lock or during final cleanup in rcu callback
+        */
+       cfg->filter = rcu_dereference_protected(prog->filter, 1);
 
        cfg->bpf_ops = prog->bpf_ops;
        cfg->bpf_name = prog->bpf_name;
@@ -294,7 +295,7 @@ static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 
        if (!tcf_hash_check(parm->index, act, bind)) {
                ret = tcf_hash_create(parm->index, est, act,
-                                     sizeof(*prog), bind, false);
+                                     sizeof(*prog), bind, true);
                if (ret < 0)
                        return ret;
 
@@ -325,7 +326,7 @@ static int tcf_bpf_init(struct net *net, struct nlattr *nla,
                goto out;
 
        prog = to_bpf(act);
-       spin_lock_bh(&prog->tcf_lock);
+       ASSERT_RTNL();
 
        if (res != ACT_P_CREATED)
                tcf_bpf_prog_fill_cfg(prog, &old);
@@ -339,14 +340,15 @@ static int tcf_bpf_init(struct net *net, struct nlattr *nla,
                prog->bpf_fd = cfg.bpf_fd;
 
        prog->tcf_action = parm->action;
-       prog->filter = cfg.filter;
-
-       spin_unlock_bh(&prog->tcf_lock);
+       rcu_assign_pointer(prog->filter, cfg.filter);
 
-       if (res == ACT_P_CREATED)
+       if (res == ACT_P_CREATED) {
                tcf_hash_insert(act);
-       else
+       } else {
+               /* make sure the program being replaced is no longer executing */
+               synchronize_rcu();
                tcf_bpf_cfg_cleanup(&old);
+       }
 
        return res;
 out: