Taking a full copy via skb_realloc_headroom() on every xmit is overkill
and wastes CPU time; all we actually need is to push on the qeth_hdr.
So rework the L2 OSA TX path to avoid the copy.
Minor complications arise because struct qeth_hdr must not cross a page
boundary. So add a new helper qeth_push_hdr() that catches this, and
falls back to the hdr cache that we already use for IQDs.

This change uncovered that qeth's TX completion takes rather long.
Now that we no longer free the original skb straight away and thus call
skb->destructor later than before, throughput regresses significantly.
For now, restore old behaviour by adding an explicit skb_orphan(),
and a big TODO to improve the TX completion time.
Tested-by: Nils Hoppmann <niho@de.ibm.com>
Signed-off-by: Julian Wiedmann <jwi@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>