whats wrong with using total_tx_bytes or buffer_info->bytecount?  it
contains the "bytes on the wire" value which will be slightly larger
than skb->len, but avoids warming the skb->len cacheline unnecessarily.

This makes sense.  I made the changes, but the limits computed are
much higher than with using sbk->len.  I suspect there may be a bug in
GSO path.

For instance, in the driver at:

  bytecount = ((segs - 1) * skb_headlen(skb)) + skb->len;

I see cases like:

  segs=34, skb_header_len(skb)=70, skb->len=49298 so bytecount=51608

Which seems reasonable... but, I also see things like:

  segs=45, skb_header_len(skb)=1522, skb->len=65226, so bytecount= 132194
                                                      ^^^^
Which doesn't seem right at all.  I am thinking there may be a bug
setting skb->data_len improperly.

Tom