On 12/19/2018 02:46 PM, Eric Dumazet wrote: > On Wed, Dec 19, 2018 at 2:23 PM Lorenzo Bianconi > <lorenzo.bianc...@redhat.com> wrote: >> >> Add napi_disable routine in gro_cells_destroy since starting from >> commit c42858eaf492 ("gro_cells: remove spinlock protecting receive >> queues") gro_cell_poll and gro_cells_destroy can run concurrently on >> napi_skbs list producing a kernel Oops if the tunnel interface is >> removed while gro_cell_poll is running. The following Oops has been >> triggered removing a vxlan device while the interface is receiving >> traffic >> > > This seems reasonable, I wonder why this bug has been hiding for so long... > > Acked-by: Eric Dumazet <eduma...@google.com> > Since we had another syzbot report involving this stuff [1], I took another look. Apparently vxlan code does not look at (dev->flags & IFF_UP) before injecting a packet (calling gro_cells_receive()) IP tunnels do this check properly. [1] HEAD commit: b71acb0e3721 Merge branch 'linus' of git://git.kernel.org/.. git tree: net-next console output: https://syzkaller.appspot.com/x/log.txt?x=14314bab400000 kernel config: https://syzkaller.appspot.com/x/.config?x=b03c5892bb940c76 dashboard link: https://syzkaller.appspot.com/bug?extid=6fe674089f9deb9f7726 compiler: gcc (GCC) 9.0.0 20181231 (experimental) I suspect the following fix is needed for vxlan diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c index 5209ee9aac47846367d7f469a7e69d08c030087e..7a443c251e604c41005d7d0f73832c22aed51768 100644 --- a/drivers/net/vxlan.c +++ b/drivers/net/vxlan.c @@ -1657,6 +1657,10 @@ static int vxlan_rcv(struct sock *sk, struct sk_buff *skb) goto drop; } + if (unlikely(!(vxlan->dev->flags & IFF_UP))) { + atomic_long_inc(&vxlan->dev->rx_dropped); + goto drop; + } stats = this_cpu_ptr(vxlan->dev->tstats); u64_stats_update_begin(&stats->syncp); stats->rx_packets++;