]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
virtio_net: Don't process redirected XDP frames when XDP is disabled
authorToshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Tue, 29 Jan 2019 00:45:57 +0000 (09:45 +0900)
committerDavid S. Miller <davem@davemloft.net>
Wed, 30 Jan 2019 22:02:43 +0000 (14:02 -0800)
Commit 8dcc5b0ab0ec ("virtio_net: fix ndo_xdp_xmit crash towards dev not
ready for XDP") tried to avoid access to unexpected sq while XDP is
disabled, but was not complete.

There was a small window which causes out of bounds sq access in
virtnet_xdp_xmit() while disabling XDP.

An example case of
 - curr_queue_pairs = 6 (2 for SKB and 4 for XDP)
 - online_cpu_num = xdp_queue_paris = 4
when XDP is enabled:

CPU 0                         CPU 1
(Disabling XDP)               (Processing redirected XDP frames)

                              virtnet_xdp_xmit()
virtnet_xdp_set()
 _virtnet_set_queues()
  set curr_queue_pairs (2)
                               check if rq->xdp_prog is not NULL
                               virtnet_xdp_sq(vi)
                                qp = curr_queue_pairs -
                                     xdp_queue_pairs +
                                     smp_processor_id()
                                   = 2 - 4 + 1 = -1
                                sq = &vi->sq[qp] // out of bounds access
  set xdp_queue_pairs (0)
  rq->xdp_prog = NULL

Basically we should not change curr_queue_pairs and xdp_queue_pairs
while someone can read the values. Thus, when disabling XDP, assign NULL
to rq->xdp_prog first, and wait for RCU grace period, then change
xxx_queue_pairs.
Note that we need to keep the current order when enabling XDP though.

- v2: Make rcu_assign_pointer/synchronize_net conditional instead of
      _virtnet_set_queues.

Fixes: 186b3c998c50 ("virtio-net: support XDP_REDIRECT")
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/virtio_net.c

index 669b65c232a21870eec7107fc246dc58782cc2af..cea52e47dc65f6bd3dfe54540a097fc4b270c70b 100644 (file)
@@ -2410,6 +2410,10 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                return -ENOMEM;
        }
 
+       old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
+       if (!prog && !old_prog)
+               return 0;
+
        if (prog) {
                prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
                if (IS_ERR(prog))
@@ -2424,21 +2428,30 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                }
        }
 
+       if (!prog) {
+               for (i = 0; i < vi->max_queue_pairs; i++) {
+                       rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
+                       if (i == 0)
+                               virtnet_restore_guest_offloads(vi);
+               }
+               synchronize_net();
+       }
+
        err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
        if (err)
                goto err;
        netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
        vi->xdp_queue_pairs = xdp_qp;
 
-       for (i = 0; i < vi->max_queue_pairs; i++) {
-               old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
-               rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
-               if (i == 0) {
-                       if (!old_prog)
+       if (prog) {
+               for (i = 0; i < vi->max_queue_pairs; i++) {
+                       rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
+                       if (i == 0 && !old_prog)
                                virtnet_clear_guest_offloads(vi);
-                       if (!prog)
-                               virtnet_restore_guest_offloads(vi);
                }
+       }
+
+       for (i = 0; i < vi->max_queue_pairs; i++) {
                if (old_prog)
                        bpf_prog_put(old_prog);
                if (netif_running(dev)) {
@@ -2451,6 +2464,12 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
        return 0;
 
 err:
+       if (!prog) {
+               virtnet_clear_guest_offloads(vi);
+               for (i = 0; i < vi->max_queue_pairs; i++)
+                       rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
+       }
+
        if (netif_running(dev)) {
                for (i = 0; i < vi->max_queue_pairs; i++) {
                        virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);