Re: [PATCH bpf-next] xsk: proper socket state check in xsk_poll

On 8/20/19 5:29 PM, Björn Töpel wrote:

On Tue, 20 Aug 2019 at 16:30, Daniel Borkmann [off-list ref] wrote:

quoted

On 8/20/19 12:04 PM, Björn Töpel wrote:

quoted

From: Björn Töpel <redacted>

The poll() implementation for AF_XDP sockets did not perform the
proper state checks, prior accessing the socket umem. This patch fixes
that by performing a xsk_is_bound() check.

Suggested-by: Hillf Danton <redacted>
Reported-by: syzbot+c82697e3043781e08802@syzkaller.appspotmail.com
Fixes: 77cd0d7b3f25 ("xsk: add support for need_wakeup flag in AF_XDP rings")
Signed-off-by: Björn Töpel <redacted>
---
   net/xdp/xsk.c | 14 ++++++++++++--
   1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index ee4428a892fa..08bed5e92af4 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c

@@ -356,13 +356,20 @@ static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
       return err;
   }

+static bool xsk_is_bound(struct xdp_sock *xs)
+{
+     struct net_device *dev = READ_ONCE(xs->dev);
+
+     return dev && xs->state == XSK_BOUND;
+}
+
   static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
   {
       bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
       struct sock *sk = sock->sk;
       struct xdp_sock *xs = xdp_sk(sk);

-     if (unlikely(!xs->dev))
+     if (unlikely(!xsk_is_bound(xs)))
               return -ENXIO;
       if (unlikely(!(xs->dev->flags & IFF_UP)))
               return -ENETDOWN;

@@ -383,6 +390,9 @@ static unsigned int xsk_poll(struct file *file, struct socket *sock,
       struct net_device *dev = xs->dev;
       struct xdp_umem *umem = xs->umem;

+     if (unlikely(!xsk_is_bound(xs)))
+             return mask;
+
       if (umem->need_wakeup)
               dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
                                               umem->need_wakeup);

@@ -417,7 +427,7 @@ static void xsk_unbind_dev(struct xdp_sock *xs)
   {
       struct net_device *dev = xs->dev;

-     if (!dev || xs->state != XSK_BOUND)
+     if (!xsk_is_bound(xs))
               return;

I think I'm a bit confused by your READ_ONCE() usage. ;-/ I can see why you're
using it in xsk_is_bound() above, but then at the same time all the other callbacks
like xsk_poll() or xsk_unbind_dev() above have a struct net_device *dev = xs->dev
right before the test. Could you elaborate?

Yes, now I'm confused as well! Digging deeper... I believe there are a
couple of places in xsk.c that do not have
READ_ONCE/WRITE_ONCE-correctness. Various xdp_sock members are read
lock-less outside the control plane mutex (mutex member of struct
xdp_sock). This needs some re-work. I'll look into using the newly

Right, so even in above two cases, the compiler could have refetched, e.g.
dev variable could have first been NULL, but xsk_is_bound() later returns
true.

introduced state member (with corresponding read/write barriers) for
this.

I'll cook some patch(es) that address this, but first it sounds like I
need to reread [1] two, or three times. At least. ;-)


Thanks,
Björn


[1] https://lwn.net/Articles/793253/

quoted

Thanks,
Daniel