Universe VME notes (was Re: BSP for Motorola MVME5500)

[Till Straumann]

Feng, Shuchen wrote:
> Hi  Bill,
> 
> I am near the completion for writting the BSP for the MVME5500.
> It successfully auto booted large (e.g. C++ based EPICS) VME based
> applications via the 10/100M ethernet.  My TODO list before the
> release of the souce codes are :
> 
> 1.  1GHZ ethernet.
> 2.  rtems-4.6.0  (right now it works on rtems-4.6.0pre3)
> 3.  build the Cexp (written by Till Straumann) to give the BSP
>     a VxWorks-like environment (e.g. dynamic loading)
> 4.  VME device write probe (the read probe is done) to
>     detect the absent VME devices.

FYI

Write probing has not been implemented (in libbspExt) for good reasons:

  1) Write cycles might be asynchronous - the universe chip
     supports 'posted writes'. Hence, you might get an interrupt
     flagging a bus error long after your 'probe' routine has returned.

     AFAIK, the universe doesn't provide a synchronization mechanism
     to enforce flushing the fifos.

     You would have to either disable posted writes (incurring
     a significant performance penalty) altogether or set up
     secondary VME windows with PWEN cleared. This would be more
     elegant but doesn't fit the EPICS devLib API.

     As a sidenote: posted writes also are problematic when talking
     to device registers (e.g. when clearing interrupt status bits):

             1) device asserts IRQ level
             2) universe acknowledges interrupt and retrieves vector
             3) user ISR is dispatched.
             4) user ISR clears the device's interupt status bit
             *** HOWEVER ***, the new value is only posted to the FIFO
             5) user ISR terminates.
             6) IRQ level is re-enabled by the universe driver
             ===> if the TXFIFO has not been flushed yet, the IRQ
                  level is still asserted!

  2) Some hardware (such as the MPC106 hostbride on my SVGM board)
     doesn't even bother to support raising machine check exceptions
     on failing VME writes.

It is recommended to exclusively use 'read probes' for device detection.


-- Till