TODO

* commit 7c734359d350 ("qlge: Size RX buffers based on MTU.", v2.6.33-rc1)
  introduced dead code in the receive routines, which should be rewritten
  anyways by the admission of the author himself, see the comment above
  qlge_build_rx_skb(). That function is now used exclusively to handle packets
  that underwent header splitting but it still contains code to handle non
  split cases.
* truesize accounting is incorrect (ex: a 9000B frame has skb->truesize 10280
  while containing two frags of order-1 allocations, ie. >16K)
* while in that area, using two 8k buffers to store one 9k frame is a poor
  choice of buffer size.
* in the "chain of large buffers" case, the driver uses an skb allocated with
  head room but only puts data in the frags.
* rename "rx" queues to "completion" queues. Calling tx completion queues "rx
  queues" is confusing.
* struct rx_ring is used for rx and tx completions, with some members relevant
  to one case only
* the flow control implementation in firmware is buggy (sends a flood of pause
  frames, resets the link, device and driver buffer queues become
  desynchronized), disable it by default
* the driver has a habit of using runtime checks where compile time checks are
  possible (ex. qlge_free_rx_buffers())
* reorder struct members to avoid holes if it doesn't impact performance
* use better-suited apis (ex. use pci_iomap() instead of ioremap())
* remove duplicate and useless comments
* fix weird line wrapping (all over, ex. the qlge_set_routing_reg() calls in
  qlge_set_multicast_list()).
* remove useless casts (ex. memset((void *)mac_iocb_ptr, ...))
* fix checkpatch issues