Brown paper bag time: the Rx descriptors are contiguous and EORbit only
marks the last descriptor in the array. OWNbit implicitly marks the end
of the Rx descriptors segment which is owned by the nic.

Originally done in rtl8169_start_xmit() by Realtek.

- register voodoo in rtl8169_hw_start().

- changed their timeout value from 100 to HZ to trigger rtl8169_phy_timer();
- s/TX_TIMEOUT/RTL8169_TX_TIMEOUT/ to have RTL8169_{TX/PHY}_TIMEOUT.

- phy capability settings allows lower or equal capability as suggested
  in Realtek's changes;
- I/O voodoo;
- no need to s/mdio_write/RTL8169_WRITE_GMII_REG/;
- s/rtl8169_hw_PHY_config/rtl8169_hw_phy_config/;
- rtl8169_hw_phy_config(): ad-hoc struct "phy_magic" to limit duplication
  of code (yep, the u16 -> int conversions should work as expected);
- variable renames and whitepace changes ignored.

- change of identification logic in rtl8169_init_board();
- {chip/rtl_chip}_info are merged in rtl_chip_info;
- misc style nits (lazy braces, SHOUTING MACROS from realtek converted to
  functions).

- removal of rtl8169_unmap_rx() (unneeded as for now).

- endianness is kept in a fscked state as it is in the original code
  (will be adressed in a later patch);
- buf_addr of an unmapped descriptor is always set to the same value 
  (cf rtl8169_unmap_tx_skb);
- nothing fancy, really.

- it forgot to update stats if the skb couldn't be expanded;
- it didn't free it either if the descriptor was not available;
- move the spin_unlock nearer of the exit point instead of duplicating
  it in the new branch.

- endianness is kept in a fscked state as it is in the original code
  (will be adressed in a later patch);
- rtl8169_rx_clear() walks the buffer ring and releases the allocated
  data buffers. It needs to be used in two places: 
  - rtl8169_init_ring() failure path;
  - normal device release (i.e. rtl8169_close);
- rtl8169_free_rx_skb() releases a Rx data buffer. Mostly an helper
  for rtl8169_rx_clear(). As such it must:
  - unmap the memory area;
  - release the skb;
  - prevent the ring descriptor from being used again;
- rtl8169_alloc_rx_skb() prepares a Rx data buffer for use.
  As such it must:
  - allocate an skb;
  - map the memory area;
  - reflect the changes in the ring descriptor.
  This function is balanced by rtl8169_free_rx_skb().
- rtl8169_unmap_rx() simply helps with the 80-columns limit.
- rtl8169_rx_fill() walks a given range of the buffer ring and
  try to turn any descriptor into a ready to use one. It returns the
  count of modified descriptors and exits if an allocation fails.
  It can be seen as balanced by rtl8169_rx_clear(). Motivation:
  - partially abstract the (usually big) piece of code for the refill
    logic at the end of the Rx interrupt;
  - factorize the refill logic and the initial ring setup.
- simple conversion of rtl8169_rx_interrupt() without rx_copybreak
  (will be adressed in a later patch).

- use pci_alloc_consistent() for Rx/Tx descriptors in rtl8169_open()
  (balanced by pci_free_consistent() on error path as well as in
  rtl8169_close());
- removal of the fields {Rx/Tx}DescArrays in struct rtl8169_private
  as there is no need to store a non-256 bytes aligned address any more;
- fix for rtl8169_open() leak when RxBufferRings allocation fails.
  Said allocation is pushed to rtl8169_init_ring() as part of an evil
  cunning plan.

into home.osdl.org:/home/torvalds/v2.5/linux

Spotted by Matteo Croce 

into home.osdl.org:/home/torvalds/v2.5/linux

From Martin Diehl.

From Martin Diehl.

Also, make sir-dev locking compatible with irport.
From Martin Diehl.

[ATM]: br2684 incorrectly handles frames recvd with FCS (by Alex Zeffertt <ajz@cambridgebroadband.com>)