From: Bernd Eckstein <redacted>
Date: Fri, 23 Nov 2018 13:51:26 +0100

The bug is not easily reproducable, as it may occur very infrequently
(we had machines with 20minutes heavy downloading before it occurred)
However, on a virual machine (VMWare on Windows 10 host) it occurred
pretty frequently (1-2 seconds after a speedtest was started)

dev->tx_skb mab be freed via dev_kfree_skb_irq on a callback
before it is set.

This causes the following problems:
- double free of the skb or potential memory leak
- in dmesg: 'recvmsg bug' and 'recvmsg bug 2' and eventually
  general protection fault

Example dmesg output:

 ...

The proposed patch eliminates a potential racing condition.
Before, usb_submit_urb was called and _after_ that, the skb was attached
(dev->tx_skb). So, on a callback it was possible, however unlikely that the
skb was freed before it was set. That way (because dev->tx_skb was not set
to NULL after it was freed), it could happen that a skb from a earlier
transmission was freed a second time (and the skb we should have freed did
not get freed at all)

Now we free the skb directly in ipheth_tx(). It is not passed to the
callback anymore, eliminating the posibility of a double free of the same
skb. Depending on the retval of usb_submit_urb() we use dev_kfree_skb_any()
respectively dev_consume_skb_any() to free the skb.

Signed-off-by: Oliver Zweigle <redacted>
Signed-off-by: Bernd Eckstein <redacted>

Applied and queued up for -stable, thanks.