2020-10-01, 17:58:40 +0200, Nicolas Dichtel wrote:
> Le 01/10/2020 à 09:59, Sabrina Dubroca a écrit :
> > Currently, we're not advertising link-netnsid when dumping IPv6
> > addresses, so the "ip -6 addr" command will not correctly interpret
> > the value of the IFLA_LINK attribute.
> >
> > For example, we'll get:
> > 9: macvlan0@macvlan0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 state
> > UP qlen 1000
> > <snip>
> >
> > Instead of:
> > 9: macvlan0@if9: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 state UP
> > qlen 1000 link-netns main
> > <snip>
> >
> > ndisc_ifinfo_sysctl_change calls inet6_fill_ifinfo without rcu or
> > rtnl, so I'm adding rcu_read_lock around rtnl_fill_link_netnsid.
> I don't think this is needed.
> ndisc_ifinfo_sysctl_change() takes a reference on the idev (with
> in6_dev_get(dev)).
The problem is veth's get_link_net implementation, even after my change in
patch 6:
static struct net *veth_get_link_net(const struct net_device *dev)
{
struct veth_priv *priv = netdev_priv(dev);
struct net_device *peer = rcu_dereference_rtnl(priv->peer);
return peer ? dev_net(peer) : dev_net(dev);
}
These commands:
ip link add type veth
sysctl net.ipv6.neigh.veth0.retrans_time_ms=2000
cause this splat:
[ 91.426764] =============================
[ 91.427445] WARNING: suspicious RCU usage
[ 91.428129] 5.9.0-rc6-net-00331-gae48bef8808b-dirty #266 Not tainted
[ 91.429209] -----------------------------
[ 91.433898] drivers/net/veth.c:1436 suspicious rcu_dereference_check() usage!
[ 91.435127]
other info that might help us debug this:
[ 91.436515]
rcu_scheduler_active = 2, debug_locks = 1
[ 91.437636] 1 lock held by sysctl/3718:
[ 91.438310] #0: ffff88806488c430 (sb_writers#5){.+.+}-{0:0}, at:
vfs_write+0x2a7/0x350
[ 91.439769]
stack backtrace:
[ 91.440552] CPU: 2 PID: 3718 Comm: sysctl Not tainted
5.9.0-rc6-net-00331-gae48bef8808b-dirty #266
[ 91.442132] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
ArchLinux 1.14.0-1 04/01/2014
[ 91.443742] Call Trace:
[ 91.444204] dump_stack+0x9a/0xd0
[ 91.444810] veth_get_link_net+0xa6/0xb0
[ 91.445534] rtnl_fill_link_netnsid+0xa2/0x130
[ 91.446330] ? rtnl_put_cacheinfo+0x190/0x190
[ 91.447120] ? memcpy+0x39/0x60
[ 91.447717] inet6_fill_ifinfo+0x2f7/0x480
I guess I could push the rcu_read_lock down into veth and vxcan's
handlers instead of the rcu_dereference_rtnl change in patch 6 and
adding this rcu_read_lock.
--
Sabrina
