the problem?
>
>The problem lies in the fact that reselection itself is an act initiated
>by the remote target and can occur at any time.  It is essentially an
>interrupt, occurring asynchronously to anything the CPU is currently
>doing.
>
>The bug lies in the fact that the hardware is broken.  During the time you
>are loading a command into the FIFO, a reselection can take place.  When
>this happens the esp clears the contents of the FIFO and posts an
>interrupt.
>
>There is a lag time between the period when the FIFO is cleared and when
>the hardware manages to assert the IRQ pin.  During this time the CPU is
>still oblivious to the fact that the FIFO has been cleared and is starting
>to be filled with data from the reselecting target.  The CPU continues to
>fill in the remaining bytes of its command, inadvertantly interspersing
>them with the incomming bytes from the target.  Eventually the IRQ pin is
>asserted and the reselect interrupt is processed.  When the interrupt
>handler examines the FIFO it sees a mishmashed collection of incoming and
>outgoing data.
>


This is why using DMA for the commands works, whilst hand stuffing doesn't.