← Back to Technotes

#21: DMA Compatibility for Expansion RAM

Author: Jim Merritt (rev. Glenn A. Baxter)
Year: 1987

... discusses the Apple IIGS Extended Memory Slot specification.

View raw text file

Apple II
Technical Notes
_____________________________________________________________________________
                                                  Developer Technical Support


Apple IIGS
#21:    DMA Compatibility for Expansion RAM

Revised by:    Glenn A. Baxter                                  November 1988
Written by:    Jim Merritt                                        August 1987

This Technical Note discusses the Apple IIGS Extended Memory Slot 
specification.
_____________________________________________________________________________

The Apple IIGS Extended Memory Slot specification provides for DMA access to 
no more than four rows of RAM on a single board through the CROW0 and CROW1 
signals.  Expansion board designs that involve more than four rows of RAM are 
not compatible with DMA accesses.  Each of the four rows can hold either 256K 
or 1 MB of data.  The design of the Fast Processor Interface (FPI) imposes 
this limit.  Each row can be organized in any of the following configurations 
to yield the respective board capacities assuming there are no more than four 
rows:

                   Chips    Configuration    Board Capacity
                     8      256K x 1 DRAM        1 MB
                     8      1 MB x 1 DRAM        4 MB
                     2      256K x 4 DRAM        1 MB
                     2      1 MB x 4 DRAM        4 MB

The CROW0 and CROW1 signals properly decode the row addresses for both normal 
and DMA cycles.  The Extended Memory Slot interface does not support the 
latching of bank address information off the data bus during a DMA cycle, and 
a card which attempts to latch the bank address will likely get the last CPU 
cycle's bank address.  Getting the last address is not a problem if it 
accidently happens to be the bank to which you wish to talk, but this is 
rarely the case.  The card gets the last CPU cycle's bank address because DMA 
essentially shuts off the CPU, so it cannot emit the bank address.  The FPI, 
which contains the DMA bank address register ($C037), does not emit the DMA 
bank address either, thus preventing bus contention with the processor as it 
is being removed from that bus.  The DMA bank address register inside the FPI 
affects the addressing and control information  that the Extended Memory Slot 
sees; it does not affect the data bus.  Therefore, during DMA, the bank 
address time is filled with what is essentially random bank address 
information.  Using this random information could result in damaging the 
contents of the memory (destroying little things like the operating system).

Suppose a card were designed to latch the bank address directly from the data 
bus with the rising edge of the PH2 clock signal.  It could use the bank 
address to derive the proper RAM row address and never bother with CROW0 and 
CROW1 at all.  Directly latching the bank address would permit the card to 
accommodate any desired RAM arrangement in 64K increments, including an odd 
number of rows.  Although the technique is valid during CPU cycles, it does 
not work during DMA cycles since the FPI never emits the DMA bank address onto 
the data bus.  During DMA cycles, any card that tries to latch the bank 
address directly, instead latches the bank address that was put on the data 
bus during the last CPU cycle, which is probably the wrong value.

Currently, there does not seem to be a solution for the DMA situation.  There 
the possibility of "limited DMA compatibility."  An example of a limited-
compatibility card would be one with six banks of memory.  It's lower four 
banks are DMA compatible since they use the CROW0 and CROW1 lines, but the 
upper two banks do not work properly with DMA.  This limited approach should 
be safe, but it is not guaranteed since DMA cards are sometimes aware of the 
total system memory and may expect, quite reasonably, to have access to all 
of the memory when in fact it does not.  There are currently no "memory 
intelligent" DMA cards, but that could change at any point.  The best we can 
suggest at this time is for hardware developers to build only four-row cards 
allowing up to 4 MB of memory, which is sufficient for most current 
applications.


Further Reference
o    Apple IIGS Hardware Reference