ipv6: Fix dst_entry refcnt bugs in ip6_tunnel
authorMartin KaFai Lau <kafai@fb.com>
Tue, 15 Sep 2015 21:30:07 +0000 (14:30 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 15 Sep 2015 21:53:05 +0000 (14:53 -0700)
Problems in the current dst_entry cache in the ip6_tunnel:

1. ip6_tnl_dst_set is racy.  There is no lock to protect it:
   - One major problem is that the dst refcnt gets messed up. F.e.
     the same dst_cache can be released multiple times and then
     triggering the infamous dst refcnt < 0 warning message.
   - Another issue is the inconsistency between dst_cache and
     dst_cookie.

   It can be reproduced by adding and removing the ip6gre tunnel
   while running a super_netperf TCP_CRR test.

2. ip6_tnl_dst_get does not take the dst refcnt before returning
   the dst.

This patch:
1. Create a percpu dst_entry cache in ip6_tnl
2. Use a spinlock to protect the dst_cache operations
3. ip6_tnl_dst_get always takes the dst refcnt before returning

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/ip6_tunnel.h
net/ipv6/ip6_gre.c
net/ipv6/ip6_tunnel.c

index 979b081a47e8cdc17908c5101e0be758f6082bf6..60b4f402f78c38c83d220f5188fa9d9b1f48adfc 100644 (file)
@@ -32,6 +32,12 @@ struct __ip6_tnl_parm {
        __be32                  o_key;
 };
 
+struct ip6_tnl_dst {
+       spinlock_t lock;
+       struct dst_entry *dst;
+       u32 cookie;
+};
+
 /* IPv6 tunnel */
 struct ip6_tnl {
        struct ip6_tnl __rcu *next;     /* next tunnel in list */
@@ -39,8 +45,7 @@ struct ip6_tnl {
        struct net *net;        /* netns for packet i/o */
        struct __ip6_tnl_parm parms;    /* tunnel configuration parameters */
        struct flowi fl;        /* flowi template for xmit */
-       struct dst_entry *dst_cache;    /* cached dst */
-       u32 dst_cookie;
+       struct ip6_tnl_dst __percpu *dst_cache; /* cached dst */
 
        int err_count;
        unsigned long err_time;
@@ -61,6 +66,8 @@ struct ipv6_tlv_tnl_enc_lim {
 } __packed;
 
 struct dst_entry *ip6_tnl_dst_get(struct ip6_tnl *t);
+int ip6_tnl_dst_init(struct ip6_tnl *t);
+void ip6_tnl_dst_destroy(struct ip6_tnl *t);
 void ip6_tnl_dst_reset(struct ip6_tnl *t);
 void ip6_tnl_dst_set(struct ip6_tnl *t, struct dst_entry *dst);
 int ip6_tnl_rcv_ctl(struct ip6_tnl *t, const struct in6_addr *laddr,
index 24f5dd8f76a831e283e8097ca27e946ee2c84c9b..646512488c289fd7069d0088e415b3ac6654568e 100644 (file)
@@ -637,17 +637,17 @@ static netdev_tx_t ip6gre_xmit2(struct sk_buff *skb,
                dst = ip6_tnl_dst_get(tunnel);
 
        if (!dst) {
-               ndst = ip6_route_output(net, NULL, fl6);
+               dst = ip6_route_output(net, NULL, fl6);
 
-               if (ndst->error)
+               if (dst->error)
                        goto tx_err_link_failure;
-               ndst = xfrm_lookup(net, ndst, flowi6_to_flowi(fl6), NULL, 0);
-               if (IS_ERR(ndst)) {
-                       err = PTR_ERR(ndst);
-                       ndst = NULL;
+               dst = xfrm_lookup(net, dst, flowi6_to_flowi(fl6), NULL, 0);
+               if (IS_ERR(dst)) {
+                       err = PTR_ERR(dst);
+                       dst = NULL;
                        goto tx_err_link_failure;
                }
-               dst = ndst;
+               ndst = dst;
        }
 
        tdev = dst->dev;
@@ -702,12 +702,9 @@ static netdev_tx_t ip6gre_xmit2(struct sk_buff *skb,
                skb = new_skb;
        }
 
-       if (fl6->flowi6_mark) {
-               skb_dst_set(skb, dst);
-               ndst = NULL;
-       } else {
-               skb_dst_set_noref(skb, dst);
-       }
+       if (!fl6->flowi6_mark && ndst)
+               ip6_tnl_dst_set(tunnel, ndst);
+       skb_dst_set(skb, dst);
 
        proto = NEXTHDR_GRE;
        if (encap_limit >= 0) {
@@ -762,14 +759,12 @@ static netdev_tx_t ip6gre_xmit2(struct sk_buff *skb,
        skb_set_inner_protocol(skb, protocol);
 
        ip6tunnel_xmit(NULL, skb, dev);
-       if (ndst)
-               ip6_tnl_dst_set(tunnel, ndst);
        return 0;
 tx_err_link_failure:
        stats->tx_carrier_errors++;
        dst_link_failure(skb);
 tx_err_dst_release:
-       dst_release(ndst);
+       dst_release(dst);
        return err;
 }
 
@@ -1223,6 +1218,9 @@ static const struct net_device_ops ip6gre_netdev_ops = {
 
 static void ip6gre_dev_free(struct net_device *dev)
 {
+       struct ip6_tnl *t = netdev_priv(dev);
+
+       ip6_tnl_dst_destroy(t);
        free_percpu(dev->tstats);
        free_netdev(dev);
 }
@@ -1248,6 +1246,7 @@ static void ip6gre_tunnel_setup(struct net_device *dev)
 static int ip6gre_tunnel_init_common(struct net_device *dev)
 {
        struct ip6_tnl *tunnel;
+       int ret;
 
        tunnel = netdev_priv(dev);
 
@@ -1259,6 +1258,13 @@ static int ip6gre_tunnel_init_common(struct net_device *dev)
        if (!dev->tstats)
                return -ENOMEM;
 
+       ret = ip6_tnl_dst_init(tunnel);
+       if (ret) {
+               free_percpu(dev->tstats);
+               dev->tstats = NULL;
+               return ret;
+       }
+
        return 0;
 }
 
index 599b0b419fbc48504a9db22f77c86b8533835d22..851cf6d1eb45a76cfa9f6c04255ba0344c76ebb7 100644 (file)
@@ -126,37 +126,90 @@ static struct net_device_stats *ip6_get_stats(struct net_device *dev)
  * Locking : hash tables are protected by RCU and RTNL
  */
 
-struct dst_entry *ip6_tnl_dst_get(struct ip6_tnl *t)
+static void __ip6_tnl_per_cpu_dst_set(struct ip6_tnl_dst *idst,
+                                     struct dst_entry *dst)
 {
-       struct dst_entry *dst = t->dst_cache;
-
-       if (dst && dst->obsolete &&
-           !dst->ops->check(dst, t->dst_cookie)) {
-               t->dst_cache = NULL;
-               dst_release(dst);
-               return NULL;
+       dst_release(idst->dst);
+       if (dst) {
+               dst_hold(dst);
+               idst->cookie = rt6_get_cookie((struct rt6_info *)dst);
+       } else {
+               idst->cookie = 0;
        }
+       idst->dst = dst;
+}
+
+static void ip6_tnl_per_cpu_dst_set(struct ip6_tnl_dst *idst,
+                                   struct dst_entry *dst)
+{
 
+       spin_lock_bh(&idst->lock);
+       __ip6_tnl_per_cpu_dst_set(idst, dst);
+       spin_unlock_bh(&idst->lock);
+}
+
+struct dst_entry *ip6_tnl_dst_get(struct ip6_tnl *t)
+{
+       struct ip6_tnl_dst *idst;
+       struct dst_entry *dst;
+
+       idst = raw_cpu_ptr(t->dst_cache);
+       spin_lock_bh(&idst->lock);
+       dst = idst->dst;
+       if (dst) {
+               if (!dst->obsolete || dst->ops->check(dst, idst->cookie)) {
+                       dst_hold(idst->dst);
+               } else {
+                       __ip6_tnl_per_cpu_dst_set(idst, NULL);
+                       dst = NULL;
+               }
+       }
+       spin_unlock_bh(&idst->lock);
        return dst;
 }
 EXPORT_SYMBOL_GPL(ip6_tnl_dst_get);
 
 void ip6_tnl_dst_reset(struct ip6_tnl *t)
 {
-       dst_release(t->dst_cache);
-       t->dst_cache = NULL;
+       int i;
+
+       for_each_possible_cpu(i)
+               ip6_tnl_per_cpu_dst_set(raw_cpu_ptr(t->dst_cache), NULL);
 }
 EXPORT_SYMBOL_GPL(ip6_tnl_dst_reset);
 
 void ip6_tnl_dst_set(struct ip6_tnl *t, struct dst_entry *dst)
 {
-       struct rt6_info *rt = (struct rt6_info *) dst;
-       t->dst_cookie = rt6_get_cookie(rt);
-       dst_release(t->dst_cache);
-       t->dst_cache = dst;
+       ip6_tnl_per_cpu_dst_set(raw_cpu_ptr(t->dst_cache), dst);
+
 }
 EXPORT_SYMBOL_GPL(ip6_tnl_dst_set);
 
+void ip6_tnl_dst_destroy(struct ip6_tnl *t)
+{
+       if (!t->dst_cache)
+               return;
+
+       ip6_tnl_dst_reset(t);
+       free_percpu(t->dst_cache);
+}
+EXPORT_SYMBOL_GPL(ip6_tnl_dst_destroy);
+
+int ip6_tnl_dst_init(struct ip6_tnl *t)
+{
+       int i;
+
+       t->dst_cache = alloc_percpu(struct ip6_tnl_dst);
+       if (!t->dst_cache)
+               return -ENOMEM;
+
+       for_each_possible_cpu(i)
+               spin_lock_init(&per_cpu_ptr(t->dst_cache, i)->lock);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(ip6_tnl_dst_init);
+
 /**
  * ip6_tnl_lookup - fetch tunnel matching the end-point addresses
  *   @remote: the address of the tunnel exit-point
@@ -271,6 +324,9 @@ ip6_tnl_unlink(struct ip6_tnl_net *ip6n, struct ip6_tnl *t)
 
 static void ip6_dev_free(struct net_device *dev)
 {
+       struct ip6_tnl *t = netdev_priv(dev);
+
+       ip6_tnl_dst_destroy(t);
        free_percpu(dev->tstats);
        free_netdev(dev);
 }
@@ -1016,17 +1072,17 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
                goto tx_err_link_failure;
 
        if (!dst) {
-               ndst = ip6_route_output(net, NULL, fl6);
+               dst = ip6_route_output(net, NULL, fl6);
 
-               if (ndst->error)
+               if (dst->error)
                        goto tx_err_link_failure;
-               ndst = xfrm_lookup(net, ndst, flowi6_to_flowi(fl6), NULL, 0);
-               if (IS_ERR(ndst)) {
-                       err = PTR_ERR(ndst);
-                       ndst = NULL;
+               dst = xfrm_lookup(net, dst, flowi6_to_flowi(fl6), NULL, 0);
+               if (IS_ERR(dst)) {
+                       err = PTR_ERR(dst);
+                       dst = NULL;
                        goto tx_err_link_failure;
                }
-               dst = ndst;
+               ndst = dst;
        }
 
        tdev = dst->dev;
@@ -1072,12 +1128,11 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
                consume_skb(skb);
                skb = new_skb;
        }
-       if (fl6->flowi6_mark) {
-               skb_dst_set(skb, dst);
-               ndst = NULL;
-       } else {
-               skb_dst_set_noref(skb, dst);
-       }
+
+       if (!fl6->flowi6_mark && ndst)
+               ip6_tnl_dst_set(t, ndst);
+       skb_dst_set(skb, dst);
+
        skb->transport_header = skb->network_header;
 
        proto = fl6->flowi6_proto;
@@ -1101,14 +1156,12 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
        ipv6h->saddr = fl6->saddr;
        ipv6h->daddr = fl6->daddr;
        ip6tunnel_xmit(NULL, skb, dev);
-       if (ndst)
-               ip6_tnl_dst_set(t, ndst);
        return 0;
 tx_err_link_failure:
        stats->tx_carrier_errors++;
        dst_link_failure(skb);
 tx_err_dst_release:
-       dst_release(ndst);
+       dst_release(dst);
        return err;
 }
 
@@ -1573,12 +1626,21 @@ static inline int
 ip6_tnl_dev_init_gen(struct net_device *dev)
 {
        struct ip6_tnl *t = netdev_priv(dev);
+       int ret;
 
        t->dev = dev;
        t->net = dev_net(dev);
        dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
        if (!dev->tstats)
                return -ENOMEM;
+
+       ret = ip6_tnl_dst_init(t);
+       if (ret) {
+               free_percpu(dev->tstats);
+               dev->tstats = NULL;
+               return ret;
+       }
+
        return 0;
 }