Commit 
c0eb454034aa ("hv_netvsc: Don't ask for additional head room in the
skb") got rid of needed_headroom setting for the driver. With the change I
hit the following issue trying to use ptkgen module:
[   57.522021] kernel BUG at net/core/skbuff.c:1128!
[   57.522021] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
...
[   58.721068] Call Trace:
[   58.721068]  [<
ffffffffa0144e86>] netvsc_start_xmit+0x4c6/0x8e0 [hv_netvsc]
...
[   58.721068]  [<
ffffffffa02f87fc>] ? pktgen_finalize_skb+0x25c/0x2a0 [pktgen]
[   58.721068]  [<
ffffffff814f5760>] ? __netdev_alloc_skb+0xc0/0x100
[   58.721068]  [<
ffffffffa02f9907>] pktgen_thread_worker+0x257/0x1920 [pktgen]
Basically, we're calling skb_cow_head(skb, RNDIS_AND_PPI_SIZE) and crash on
    if (skb_shared(skb))
        BUG();
We probably need to restore needed_headroom setting (but shrunk to
RNDIS_AND_PPI_SIZE as we don't need more) to request the required headroom
space. In theory, it should not give us performance penalty.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
 
        net->ethtool_ops = ðtool_ops;
        SET_NETDEV_DEV(net, &dev->device);
 
+       /* We always need headroom for rndis header */
+       net->needed_headroom = RNDIS_AND_PPI_SIZE;
+
        /* Notify the netvsc driver of the new device */
        memset(&device_info, 0, sizeof(device_info));
        device_info.ring_size = ring_size;