Thread (32 messages) flat view 32 messages, 4 authors, 22h ago

Re: [PATCH v6 net-next 08/15] ipv4: Replace &arp_tbl with arp_table(net).

From: Kuniyuki Iwashima <kuniyu@google.com>
Date: 2026-09-04 16:16:23

On Thu, Sep 3, 2026 at 8:10 PM [off-list ref] wrote:
Thank you for your contribution! Sashiko AI review found 3 potential
issue(s) to consider:

Critical: 0 · High: 1 · Medium: 0 (1) · Low: 2

- [High] mlx5e_tc_update_neigh_used_value()
  (drivers/net/ethernet/mellanox/mlx5/core/en/tc_tun_encap.c) now…
This is a pre-existing issue as mentioned in the commit message.

  3. mlx5e_tc_update_neigh_used_value() dereferences nhe->neigh_dev
     in trace_mlx5e_tc_update_neigh_used_value().

- [Low] estill instructs switchdev
  driver authors to resolve nexthops against, and monitor updates to,…
- [Low] The patch introduces new reads of the RCU-protected netns pointer
  dev->nd_net via the non-RCU accessor dev_net() in paths that run only…
I'll follow up on these if there are no other critical issues.


Pre-existing issues:
- [Medium] The neighbour fallback path in mlx5e_ipsec_init_macs()
  (drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c, label…

--

ipv4: Replace &arp_tbl with arp_table(net).

In preparation for per-netns neighbour tables, this converts all users of
the global arp_tbl to the arp_table(net) accessor, makes arp_tbl static in
net/ipv4/arp.c and drops its EXPORT_SYMBOL. It also removes some driver
checks that compared n->tbl against &arp_tbl and &nd_tbl.

A few questions below.
quoted
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/tc_tun_encap.c b/drivers/net/ethernet/mellanox/mlx5/core/en/tc_tun_encap.c
index 8b827201935e..33737a29ab97 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/tc_tun_encap.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/tc_tun_encap.c
@@ -393,20 +393,10 @@ void mlx5e_tc_update_neigh_used_value(struct mlx5e_neigh_hash_entry *nhe)
      struct mlx5e_encap_entry *e = NULL;
      struct mlx5e_tc_flow *flow;
      struct mlx5_fc *counter;
-     struct neigh_table *tbl;
      bool neigh_used = false;
      struct neighbour *n;
      u64 lastuse;

-     if (m_neigh->family == AF_INET)
-             tbl = &arp_tbl;
-#if IS_ENABLED(CONFIG_IPV6)
-     else if (m_neigh->family == AF_INET6)
-             tbl = &nd_tbl;
-#endif
-     else
-             return;
-
      /* mlx5e_get_next_valid_encap() releases previous encap before returning
       * next one.
       */
@@ -447,12 +437,23 @@ void mlx5e_tc_update_neigh_used_value(struct mlx5e_neigh_hash_entry *nhe)
      trace_mlx5e_tc_update_neigh_used_value(nhe, neigh_used);

      if (neigh_used) {
+             struct net_device *dev = READ_ONCE(nhe->neigh_dev);
+             struct net *net = dev_net(dev);
+             struct neigh_table *tbl;
[Severity: High]
Is nhe->neigh_dev safe to dereference here?  As far as I can tell no
reference is taken on it.  It is only stored:

drivers/net/ethernet/mellanox/mlx5/core/en/rep/neigh.c:mlx5e_rep_neigh_entry_create() {
        ...
        WRITE_ONCE((*nhe)->neigh_dev, neigh_dev);
        ...
}

and there is no netdev_hold()/dev_hold() for it anywhere, and no
NETDEV_UNREGISTER handling in the mlx5 rep neigh code that would clear it.

The rest of the driver only ever compares the field, never dereferences it:

drivers/net/ethernet/mellanox/mlx5/core/en/rep/neigh.c:mlx5e_rep_neigh_update() {
        ...
        same_dev = READ_ONCE(nhe->neigh_dev) == n->dev;
        ...
}

Before this patch the pointer was only handed to neigh_lookup(), which
never loads from the netdev (arp_hashfn() uses hash32_ptr(dev) and the
lookup only compares n->dev == dev), so a stale pointer was harmless.
Now dev_net(dev) reads dev->nd_net and the result is dereferenced again by
arp_table(net) as net->neigh_tables[NEIGH_ARP_TABLE].

If the cached route egress device (for example a vlan or bond on top of the
uplink) was unregistered and freed while the nhe is still alive, does this
read freed memory and then index a garbage struct net?  The reachable path
is the periodic work:

drivers/net/ethernet/mellanox/mlx5/core/en/rep/neigh.c:mlx5e_rep_neigh_stats_work() {
        ...
        while ((nhe = mlx5e_get_next_nhe(rpriv, nhe)) != NULL)
                mlx5e_tc_update_neigh_used_value(nhe);
        ...
}

and rtnl_lock() held there does not keep an already-unregistered netdev
alive, since free_netdev() runs from netdev_run_todo() after rtnl_unlock().

Also, note 3 in the changelog says:

  "3. mlx5e_tc_update_neigh_used_value() dereferences nhe->neigh_dev
      in trace_mlx5e_tc_update_neigh_used_value()."

Does that justification hold?  The tracepoint dereference lives in
TP_fast_assign():

drivers/net/ethernet/mellanox/mlx5/core/diag/en_tc_tracepoint.h:
        TP_STRUCT__entry(__string(devname, nhe->neigh_dev->name)

which only runs when the tracepoint is enabled, so it does not show the
pointer is safe to dereference unconditionally.

Would deriving the netns from a refcounted device such as e->out_dev or
priv->netdev work instead?
quoted
              nhe->reported_lastuse = jiffies;

+#if IS_ENABLED(CONFIG_IPV6)
+             if (m_neigh->family != AF_INET)
+                     tbl = &nd_tbl;
+             else
+#endif
+                     tbl = arp_table(net);
+
              /* find the relevant neigh according to the cached device and
               * dst ip pair
               */
-             n = neigh_lookup(tbl, &m_neigh->dst_ip, READ_ONCE(nhe->neigh_dev));
+             n = neigh_lookup(tbl, &m_neigh->dst_ip, dev);
              if (!n)
                      return;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
index db260e3d1412..37a8ddee3ea1 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
@@ -364,9 +365,10 @@ static void mlx5e_ipsec_init_macs(struct mlx5e_ipsec_sa_entry *sa_entry,
      return;

 neigh:
-     n = neigh_lookup(&arp_tbl, pkey, netdev);
+     tbl = arp_table(dev_net(netdev));
+     n = neigh_lookup(tbl, pkey, netdev);
      if (!n) {
-             n = neigh_create(&arp_tbl, pkey, netdev);
+             n = neigh_create(tbl, pkey, netdev);
              if (IS_ERR(n))
                      return;
              neigh_event_send(n, NULL);
[Severity: Medium]
This isn't a bug introduced by this patch, since the pre-patch code already
used &arp_tbl unconditionally here, but should this fallback pick the table
based on addrs->family?

The AF_INET6 branch above can reach this label with a 16-byte key:

drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c:mlx5e_ipsec_init_macs() {
        ...
                        pkey = &addrs->saddr.a6;
        ...
        case AF_INET6:
                if (!IS_ENABLED(CONFIG_IPV6) ||
                    ip6_dst_lookup(dev_net(netdev), NULL, &rt_dst_entry, &fl6))
                        goto neigh;
        ...
}

and the ARP table has .family = AF_INET and .key_len = 4, so the lookup
hashes and compares only the first 4 bytes of the IPv6 address.  Can
neigh_create() then install a bogus ARP entry and send an ARP request for
an address made from those 4 bytes, leaving the IPv6 packet-offload SA with
attrs->drop set or an unrelated MAC?  A separate fix using the netns ND
table for AF_INET6 seems needed.
quoted
diff --git a/drivers/net/ethernet/netronome/nfp/flower/tunnel_conf.c b/drivers/net/ethernet/netronome/nfp/flower/tunnel_conf.c
index ca30702f8878..551db6417405 100644
--- a/drivers/net/ethernet/netronome/nfp/flower/tunnel_conf.c
+++ b/drivers/net/ethernet/netronome/nfp/flower/tunnel_conf.c
@@ -235,7 +236,8 @@ void nfp_tunnel_keep_alive(struct nfp_app *app, struct sk_buff *skb)
              if (!netdev)
                      continue;

-             n = neigh_lookup(&arp_tbl, &ipv4_addr, netdev);
+             tbl = arp_table(dev_net(netdev));
+             n = neigh_lookup(tbl, &ipv4_addr, netdev);
              if (!n)
                      continue;
[ ... ]
quoted
diff --git a/drivers/net/vxlan/vxlan_core.c b/drivers/net/vxlan/vxlan_core.c
index 459f19f7071e..03e932744e3e 100644
--- a/drivers/net/vxlan/vxlan_core.c
+++ b/drivers/net/vxlan/vxlan_core.c
@@ -1841,11 +1841,12 @@ static int vxlan_err_lookup(struct sock *sk, struct sk_buff *skb)

 static int arp_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni)
 {
+     struct neigh_table *tbl = arp_table(dev_net(dev));
      struct vxlan_dev *vxlan = netdev_priv(dev);
+     struct neighbour *n;
      struct arphdr *parp;
      u8 *arpptr, *sha;
      __be32 sip, tip;
-     struct neighbour *n;
[ ... ]
quoted
diff --git a/include/net/arp.h b/include/net/arp.h
index f8d18b1f8b28..e932def63d62 100644
--- a/include/net/arp.h
+++ b/include/net/arp.h
@@ -7,8 +7,6 @@
 #include <linux/hash.h>
 #include <net/neighbour.h>

-extern struct neigh_table arp_tbl;
-
 static inline struct neigh_table *arp_table(struct net *net)
 {
      return net->neigh_tables[NEIGH_ARP_TABLE];
@@ -25,10 +23,12 @@ static inline u32 arp_hashfn(const void *pkey, const struct net_device *dev, u32
 #ifdef CONFIG_INET
 static inline struct neighbour *__ipv4_neigh_lookup_noref(struct net_device *dev, u32 key)
 {
+     struct neigh_table *tbl = arp_table(dev_net(dev));
+
      if (dev->flags & (IFF_LOOPBACK | IFF_POINTOPOINT))
              key = INADDR_ANY;

-     return ___neigh_lookup_noref(&arp_tbl, neigh_key_eq32, arp_hashfn, &key, dev);
+     return ___neigh_lookup_noref(tbl, neigh_key_eq32, arp_hashfn, &key, dev);
 }
[Severity: Low]
Should these new netns reads use dev_net_rcu() instead of dev_net()?

possible_net_t holds a struct net __rcu *net published with
rcu_assign_pointer() via write_pnet(), and the two accessors differ:

include/net/net_namespace.h:
        static inline struct net *read_pnet(const possible_net_t *pnet)
        {
                return rcu_dereference_protected(pnet->net, true);
        }
        ...
        static inline struct net *read_pnet_rcu(const possible_net_t *pnet)
        {
                return rcu_dereference(pnet->net);
        }

dev_net() goes through read_pnet(), which asserts unconditionally that the
caller holds RTNL or a device reference, so it silently turns off
lockdep/sparse RCU checking on these new dereferences.

The sites added by this patch that run under rcu_read_lock() only are
__ipv4_neigh_lookup_noref() and ip_neigh_gw4() (reached from
ip_finish_output2()), br_do_proxy_suppress_arp(), the vxlan
arp_reduce()/route_shortcircuit() xmit paths, and nfp_tunnel_keep_alive(),
which is called inside its own rcu_read_lock().  None of them held RTNL or
a netns reference before, and none of them read dev->nd_net at all before
this patch.

The dev_net() uses in inetdev_init(), inetdev_destroy() and arp_ifdown()
look fine since RTNL is held there.
quoted
diff --git a/include/net/route.h b/include/net/route.h
index 45290177a33c..6b55de2e4df8 100644
--- a/include/net/route.h
+++ b/include/net/route.h
@@ -405,8 +405,11 @@ static inline struct neighbour *ip_neigh_gw4(struct net_device *dev,
      struct neighbour *neigh;

      neigh = __ipv4_neigh_lookup_noref(dev, (__force u32)daddr);
-     if (unlikely(!neigh))
-             neigh = __neigh_create(&arp_tbl, &daddr, dev, false);
+     if (unlikely(!neigh)) {
+             struct neigh_table *tbl = arp_table(dev_net(dev));
+
+             neigh = __neigh_create(tbl, &daddr, dev, false);
+     }

      return neigh;
 }
diff --git a/net/bridge/br_arp_nd_proxy.c b/net/bridge/br_arp_nd_proxy.c
index 87a40e2f1c50..92a3abb1e357 100644
--- a/net/bridge/br_arp_nd_proxy.c
+++ b/net/bridge/br_arp_nd_proxy.c
@@ -193,7 +193,7 @@ void br_do_proxy_suppress_arp(struct sk_buff *skb, struct net_bridge *br,
              return;
      }

-     n = neigh_lookup(&arp_tbl, &tip, vlandev);
+     n = neigh_lookup(arp_table(dev_net(vlandev)), &tip, vlandev);
      if (n) {
              struct net_bridge_fdb_entry *f;
              u8 ha[ETH_ALEN] __aligned(2);
diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index 824c65c70408..a44fa68fdd07 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -149,7 +149,7 @@ static const struct neigh_ops arp_direct_ops = {
      .connected_output =     neigh_direct_output,
 };

-struct neigh_table arp_tbl = {
+static struct neigh_table arp_tbl = {
      .family         = AF_INET,
      .key_len        = 4,
      .protocol       = cpu_to_be16(ETH_P_IP),
@@ -182,7 +182,6 @@ struct neigh_table arp_tbl = {
      .gc_thresh2     = 512,
      .gc_thresh3     = 1024,
 };
-EXPORT_SYMBOL(arp_tbl);
[Severity: Low]
Should Documentation/networking/switchdev.rst be updated in this patch too?

Its Nexthop Resolution section still tells switchdev driver authors to use
the symbol that becomes file-local here, in four places:

  "address discovery comes via the ARP (or ND) process and is available via
   the arp_tbl neighbor table."

  "The driver can monitor for updates to arp_tbl using the netevent notifier
   NETEVENT_NEIGH_UPDATE. ... The driver implements ndo_neigh_destroy to know
   when arp_tbl neighbor entries are purged from the port."

At this commit, git grep arp_tbl matches only net/ipv4/arp.c and
Documentation/networking/switchdev.rst, and the doc points at rocker as the
reference implementation while rocker itself is converted to
arp_table(&init_net) here.  The doc also never mentions arp_table(net) or
the new requirement to pick the table belonging to the device's netns (or
init_net for netns-immutable devices).  The text is still unchanged at the
end of the series (d4be111038a59be6217f9b018ec033041ab7bf2a).

[ ... ]

--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260902203722.926528-1-kuniyu%40google.com
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help