netfilter: ipt_CLUSTERIP: two more fixes
Subsystems (2)
Florian Westphal
maintainercover netfilter: ipt_CLUSTERIP: two more fixes
During code audit I found two more bugs in CLUSTERIP, first one is a refcount leak, second is possible use-after free due to kfree() of rcu-protected (and already visible) structure.
I think its time to remove this target, so, if anyone has a good reason to not nuke it in nf-next please let me know soon.
Florian Westphal
maintainerPATCH 1/2 netfilter: ipt_CLUSTERIP: put config struct if we can't increment ct refcount
This needs to put() the entry to avoid a resource leak in error path.
--- net/ipv4/netfilter/ipt_CLUSTERIP.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv4/netfilter/ipt_CLUSTERIP.c +6 -2
@@ -496,12 +496,15 @@ static int clusterip_tg_check(const struct xt_tgchk_param *par)return PTR_ERR (config); } } - cipinfo->config = config; ret =nf_ct_netns_get (par->net, par->family); -if (ret <0 ) +if (ret <0 ) {pr_info ("cannot load conntrack support for proto=%u\n" , par->family); +clusterip_config_entry_put (par->net, config); +clusterip_config_put (config); +return ret; + }if (!par->net->xt.clusterip_deprecated_warning) { pr_info("ipt_CLUSTERIP is deprecated and it will removed soon, "
@@ -509,6 +512,7 @@ static int clusterip_tg_check(const struct xt_tgchk_param *par) par->net->xt.clusterip_deprecated_warning = true; } + cipinfo->config = config;return ret; }
Signature · 1 line
2.16.1
Florian Westphal
maintainerPATCH 2/2 netfilter: ipt_CLUSTERIP: put config instead of freeing it
In reply to Florian WestphalOnce struct is added to per-netns list it becomes visible to other cpus, so we cannot use kfree().
Also delay setting entries refcount to 1 until after everything is initialised so that when we call clusterip_config_put() in this spot entries is still zero.
--- net/ipv4/netfilter/ipt_CLUSTERIP.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
net/ipv4/netfilter/ipt_CLUSTERIP.c +4 -3
@@ -232,7 +232,6 @@ clusterip_config_init(struct net *net, const struct ipt_clusterip_tgt_info *i, c->hash_mode = i->hash_mode; c->hash_initval = i->hash_initval;refcount_set (&c->refcount,1 ); -refcount_set (&c->entries,1 );spin_lock_bh (&cn->lock);if (__clusterip_config_find (net, ip)) {
@@ -263,8 +262,10 @@ clusterip_config_init(struct net *net, const struct ipt_clusterip_tgt_info *i, c->notifier.notifier_call = clusterip_netdev_event; err =register_netdevice_notifier (&c->notifier); -if (!err) +if (!err) { +refcount_set (&c->entries,1 );return c; + }#ifdef CONFIG_PROC_FS proc_remove (c->pde);
@@ -273,7 +274,7 @@ clusterip_config_init(struct net *net, const struct ipt_clusterip_tgt_info *i,spin_lock_bh (&cn->lock);list_del_rcu (&c->list);spin_unlock_bh (&cn->lock); -kfree (c); +clusterip_config_put (c);return ERR_PTR (err); }
Signature · 1 line
2.16.1
Pablo Neira Ayuso
maintainerQuoted 6 lines · Florian Westphal
During code audit I found two more bugs in CLUSTERIP, first one is a refcount leak, second is possible use-after free due to
kfree()of rcu-protected (and already visible) structure.I think its time to remove this target, so, if anyone has a good reason to not nuke it in nf-next please let me know soon.
We have the cluster match replacement that has been around for ~10 years, it just needs an extra arptable rules.
So I'm fine with this nuke, actually I remember we talked about deprecating it.