cvs commit: src/sys/dev/netif/ral rt2661.c rt2661var.h

[Sepherosa Ziehau <sephe@xxxxxxxxxxxxxxxxxxxxxxx>]

sephe       2006/11/20 07:03:26 PST

DragonFly src repository

  Modified files:
    sys/dev/netif/ral    rt2661.c rt2661var.h 
  Log:
  Rework rt2661 TX interrupts processing
  
  The Clues: [1]
  "TX DMA Done Interrupt":
  A frame has been DMAed to ASIC, probably stored in an ASIC internal
  TX FIFO.  The TX descriptor and the resource in host memory are useless
  to ASIC
  
  "TX Done Interrupt":
  The transfering of a frame, which probably is from ASIC internal TX
  FIFO, is done by ASIC, and the information concerning this transfering
  is available
  
  The Problems:
  - The original driver reduces TX ring occupation counter on "TX Done
    Interrupt", though TX descriptors and related resources are already
    reaped on "TX DMA Done Interrupt"
  - For some (maybe all ;-) ASIC, "TX Done Interrupt" comes less frequent
    than "TX DMA Done Interrupt", probably due to the "ASIC internal TX
    FIFO", or just because it is not intended to be used in the original
    driver's way ;-)
  
  So useless TX descriptors slowly occupy TX ring and result in an almost
  full TX ring at the most of the time if new frames keep coming at high
  rate (e.g. during a long time scp(1))
  
  The Fixes:
  - Move rate control context out of rt2661_tx_data, so it is not 1:1 mapping
    with TX descriptor.  This degenerates rt2661_tx_data to rt2661_rx_data,
    so merge them and rename the new structure rt2661_data
  - Add a rate control context queue, which temporarily holds rate control
    contexts for each data transfering
  - Allocate the rate control context on fly during rt2661_tx_data().  Since
    this context is allocated using M_NOWAIT, it is possible that the
    alloctation fails.  if this happens, later rate control information will
    not be gathered (I think this is acceptable :-) and the reference to the
    ieee80211_node is freed immediately at the end of rt2661_tx_data().
  
  After above three changes reaping of TX descriptors and related resources no
  longer needs to be synchronized with rate control information gathering.
  
  - Change the meaning of rt2661_tx_desc.qid, it now indicates whether rate
    control information needs to be gathered (it is not necessary at all for
    control/management frames)
  - On "TX DMA Done Interrupt" instead of on "TX Done Interrupt", if some TX
    descriptors and related resources are reaped, reduce TX ring occupation
    counter, and kick ifnet.if_start
  
  As of this commit, my Edimax EW-7608PG can tranfer data stably [2] without
  being choked half way.
  
  #
  # [1] These clues are infered from a comment concering TX descriptor
  #     structure in Linux RT61 driver
  # [2] Though data tranfering is quite stable, its performance is still
  #     far from expectation.  In infrastructured mode, 3 meters away from
  #     the AP, I only got 6Mbits/s - 7Mbits/s, but under the same conditions
  #     receiving performance is quite OK (~20Mbits/s)
  #
  
  Revision  Changes    Path
  1.8       +107 -81   src/sys/dev/netif/ral/rt2661.c
  1.3       +11 -11    src/sys/dev/netif/ral/rt2661var.h


http://www.dragonflybsd.org/cvsweb/src/sys/dev/netif/ral/rt2661.c.diff?r1=1.7&r2=1.8&f=u
http://www.dragonflybsd.org/cvsweb/src/sys/dev/netif/ral/rt2661var.h.diff?r1=1.2&r2=1.3&f=u