David Wilder <dwilder@xxxxxxxxxx> wrote:
> This crash happened on a ppc64le system running ltp network tests when ltp script ran "rmmod iptable_mangle".
>
> [213425.602369] BUG: Kernel NULL pointer dereference at 0x00000010
> [213425.602388] Faulting instruction address: 0xc008000000550bdc
[..]
> In the crash we find in iptable_mangle_hook() that state->net->ipv4.iptable_mangle=NULL causing a NULL pointer dereference. net->ipv4.iptable_mangle is set to NULL in iptable_mangle_net_exit() and called when ip_mangle modules is unloaded. A rmmod task was found in the crash dump. A 2nd crash showed the same problem when running "rmmod iptable_filter" (net->ipv4.iptable_filter=NULL).
>
> Once a hook is registered packets will picked up a pointer from: net->ipv4.iptable_$table. The patch adds a call to synchronize_net() in ipt_unregister_table() to insure no packets are in flight that have picked up the pointer before completing the un-register.
>
> This change has has prevented the problem in our testing. However, we have concerns with this change as it would mean that on netns cleanup, we would need one synchronize_net() call for every table in use. Also, on module unload, there would be one synchronize_net() for every existing netns.
Yes, I agree with the analysis.
> Signed-off-by: David Wilder <dwilder@xxxxxxxxxx>
> ---
> net/ipv4/netfilter/ip_tables.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/net/ipv4/netfilter/ip_tables.c b/net/ipv4/netfilter/ip_tables.c
> index c2670ea..97c4121 100644
> --- a/net/ipv4/netfilter/ip_tables.c
> +++ b/net/ipv4/netfilter/ip_tables.c
> @@ -1800,8 +1800,10 @@ int ipt_register_table(struct net *net, const struct xt_table *table,
> void ipt_unregister_table(struct net *net, struct xt_table *table,
> const struct nf_hook_ops *ops)
> {
> - if (ops)
> + if (ops) {
> nf_unregister_net_hooks(net, ops, hweight32(table->valid_hooks));
> + synchronize_net();
> + }
I'd wager ebtables, arptables and ip6tables have the same bug.
The extra synchronize_net() isn't ideal. We could probably do it this
way and then improve in a second patch.
One way to fix this without a new synchronize_net() is to switch all
iptable_foo.c to use ".pre_exit" hook as well.
pre_exit would unregister the underlying hook and .exit would to the
table freeing.
Since the netns core already does an unconditional synchronize_rcu after
the pre_exit hooks this would avoid the problem as well.
