The 8168 ignores the requests to fetch the Tx descriptors when
the relevant TxPoll bit is already set. It easily kills the
performances of the 8168. David Gundersen has noticed that it
is enough to wait for the completion of the DMA transfer (NPQ
bit is cleared) before writing the TxPoll register again.

The extra IO traffic added by the proposed workaround could be
minimalized but it is not a high-priority task.

Fix for:
http://bugzilla.kernel.org/show_bug.cgi?id=7924
http://bugzilla.kernel.org/show_bug.cgi?id=8688
(http://bugzilla.kernel.org/show_bug.cgi?id=7555 ?)
Signed-off-by: Francois Romieu <romieu@fr.zoreil.com>
Cc: David Gundersen <gundy@iinet.net.au>
Cc: Edward Hsu <edward_hsu@realtek.com.tw>