- in xl_attach(), print xl card type ("90X"/"90XB"), if verbose booting is

- in xl_attach(), print xl card type ("90X"/"90XB"), if verbose booting is
  enabled
- in xl_start() and xl_start_90xB(), do BUS_DMASYNC_PREWRITE after *all* of
  the CPU memory updating is finished
- use UpPoll feature for XL_TYPE_905B typed card.  this can eliminate device
  register accessing during xl_rxeof() thus improving recv performance(*) with
  less CPU overhead
  Inspired-by: FreeBSD pci/if_xl.c rev1.58's commit log
(*)
benchmark setting:
1) sephe:xl0 (dev-1.3.6 Sep.24) <----> sephe-test:vr0 (dev-1.3.6 Sep.24)
2) except for the different xl driver, rest part of the kernel running on
   sephe is same
3) xl0 is a <3Com 3c905C-TX Fast Etherlink XL> card, XL_TYPE_905B typed
4) netperf is executed with only "-H <host> -p <port> -t <test type>"

benchmark result:
1) run netperf from sephe-test:vr0
   17.0% (41.30[new xl]/35.28[old xl]) performance boost is observed for
   "TCP STREAM TEST"
2) run netperf from sephe:xl0
   13.6% (7405.79[new xl]/6518.39[old xl]) performance boost is observed for
   "TCP REQUEST/RESPONSE TEST"
3) minor performance boosts are observed in the rest of netperf "TEST"s

# Thank Bill Paul for that detailed commit log