Here's a patch for -mm for now.  Not sure whose territory this falls
  in, so I'm sending it to everyone I can think of. :)
  
  Some time ago I did some experiments with using PPP multilink over
  largish numbers of channels (up to 32).  The TCP performance was
  woeful due to wildly fluctuating packet latencies, which turned out to
  be because we would sometimes split a packet across all 32 channels,
  and sometimes we would send a whole packet down a single channel.
  
  This patch fixes those problems by being a bit cleverer about how the
  packets are split across the available channels, and in particular, it
  waits until at least half of the channels can take another fragment
  before starting to split up the next packet.
  
  The patch also fixes a buglet in the multilink reconstruction code
  where it would discard incoming packets that had just the multilink
  header and no data.  Such packets are valid and shouldn't be
  discarded.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>