                           PowerPC 440

                    Last Update: September 11, 2002
=======================================================================


OVERVIEW
============

Support for the ppc440 is contained in the cpu/ppc44x directory
and enabled via the CONFIG_440 flag. It is largely based on the
405gp code. A sample board support implementation is contained
in the board/ebony directory.

All testing was performed using the IBM Ebony board using both
Rev B and Rev C silicon. However, since the Rev B. silicon has
extensive errata, support for Rev B. is minimal (it boots, and
features such as i2c, pci, tftpboot, etc. seem to work ok).
The expectation is that all new board designs will be using
Rev C or later parts -- if not, you may be in for a rough ride ;-)

The ppc440 port does a fair job of keeping "board-specific" code
out of the "cpu-specific" source. The goal of course was to
provide mechanisms for each board to customize without having
to clutter the cpu-specific source with a lot of ifdefs. Most
of these mechanisms are described in the following sections.


MEMORY MANAGEMENT
=================

The ppc440 doesn't run in "real mode". The MMU must be active
at all times. Additionally, the 440 implements a 36-bit physical
memory space that gets mapped into the PowerPC 32-bit virtual
address space. So things like memory-mapped peripherals, etc must
all be mapped in. Once this is done, the 32-bit virtual address
space is then viewed as though it were physical memory.

However, this means that memory, peripherals, etc can be configured
to appear (mostly) anywhere in the virtual address space. Each board
must define its own mappings using the tlbtab (see board/ebony/init.S).
The actual TLB setup is performed by the cpu-specific code.

Although each board is free to define its own mappings, there are
several definitions to be aware of. These definitions may be used in
the cpu-specific code (vs. board-specific code), so you should
at least review these before deciding to make any changes ... it
will probably save you some headaches ;-)

CFG_SDRAM_BASE - The virtual address where SDRAM is mapped (always 0)

CFG_FLASH_BASE - The virtual address where FLASH is mapped.

CFG_PCI_MEMBASE - The virtual address where PCI-bus memory is mapped.
    This mapping provides access to PCI-bus memory.

CFG_PERIPHERAL_BASE - The virtual address where the 440 memory-mapped
    peripherals are mapped. (e.g. -- UART registers, IIC registers, etc).

CFG_ISRAM_BASE - The virtual address where the 440 internal SRAM is
    mapped. The internal SRAM is equivalent to 405gp OCM and is used
    for the initial stack.

CFG_PCI_BASE - The virtual address where the 440 PCI-x bridge config
    registers are mapped.

CFG_PCI_TARGBASE - The PCI address that is mapped to the virtual address
    defined by CFG_PCI_MEMBASE.


UART / SERIAL
=================

The UART port works fine when an external serial clock is provided
(like the one on the Ebony board) and when using internal clocking.
This is controlled with the CFG_EXT_SERIAL_CLOCK flag. When using
internal clocking, the "ideal baud rate" settings in the 440GP
user manual are automatically calculated.

CONFIG_SERIAL_SOFTWARE_FIFO enables interrupt-driven serial operation.
But the last time I checked, interrupts were initialized after the
serial port causing the interrupt handler to be removed from the
handler table. This will probably be fixed soon ... or fix it
yourself and submit a patch :-)


I2C
=================

The i2c utilities have been tested on both Rev B. and Rev C. and
look good. The iprobe command implementation has been updated to
allow for 'skipped' addresses. Some i2c slaves are write only and
cause problems when a probe (read) is performed (for example the
CDCV850 clock controller at address 0x69 on the ebony board).

To prevent probing certain addresses you can define the
CFG_I2C_NOPROBES macro in your board-specific header file. When
defined, all specified addresses are skipped during a probe.
The addresses that are skipped will be displayed in the output
of the iprobe command.

For example, to prevent probing address 0x69, define the macro as
follows:

#define CFG_I2C_NOPROBES {0x69}

Similarly, to prevent probing addresses 0x69 and 0x70, define the
macro a:

#define CFG_I2C_NOPROBES {0x69, 0x70}


DDR SDRAM CONTROLLER
====================

SDRAM controller intialization using Serial Presence Detect (SPD) is
now supported (thanks Jun). It is enabled by defining CONFIG_SPD_EEPROM.
The i2c eeprom addresses are controlled by the SPD_EEPROM_ADDRESS macro.

NOTE: The SPD_EEPROM_ADDRESS macro is defined differently than for other
processors. Traditionally, it defined a single address. For the 440 it
defines an array of addresses to support multiple banks. Address order
is significant: the addresses are used in order to program the BankN
registers. For example, two banks with i2c addresses of 0x53 (bank 0)
and 0x52 (bank 1) would be defined as follows:

#define SPD_EEPROM_ADDRESS {0x53,0x52}


PCI-X BRIDGE
====================

PCI is an area that requires lots of flexibility since every board has
its own set of constraints and configuration. This section describes the
440 implementation.

CPC0_STRP1[PISE] -- if the PISE strap bit is not asserted, PCI init
is aborted and an indication is printed. This is NOT considered an
error -- only an indication that PCI shouldn't be initialized. This
gives you a chance to edit the i2c bootstrap eeproms using the i2c
utilities once you get to the U-Boot command prompt. NOTE: the default
440 bootstrap options (not using i2c eeprom) negates this bit.

The cpu-specific code sets up a default pci_controller structure
that maps in a single PCI I/O space and PCI memory space. The I/O
space begins at PCI I/O address 0 and the PCI memory space is
256 MB starting at PCI address CFG_PCI_TARGBASE. After the
pci_controller structure is initialized, the cpu-specific code will
call the routine pci_pre_init() if the CFG_PCI_PRE_INIT flag is
defined. This routine is implemented by board-specific code & is where
the board can over-ride/extend the default pci_controller structure
settings and do other pre-initialization tasks. If pci_pre_init()
returns a value of zero, PCI initialization is aborted; otherwise the
controller structure is registered and initialization continues.

The default 440GP PCI target configuration is minimal -- it assumes that
the strapping registers are set as necessary. Since the strapping bits
provide very limited flexibility, you may want to customize the boards
target configuration. If CFG_PCI_TARGET_INIT is defined, the cpu-specific
code will call the routine pci_target_init() which you must implement
in your board-specific code.

Target initialization is completed by the cpu-specific code by
initializing the subsystem id and subsystem vendor id, and then ensuring
that the 'enable host configuration' bit in the PCIX0_BRDGOPT2 is set.

The default PCI master initialization maps in 256 MB of pci memory
starting at PCI address CFG_PCI_MEMBASE. To customize this, define
PCI_MASTER_INIT. This will call the routine pci_master_init() in your
board-specific code rather than performing the default master
initialization.

The decision to perform PCI host configuration must often be determined
at run time. The ppc440 port differs from most other implementations in
that it requires the board to determine its host configuration at run
time rather than by using compile-time flags. This shouldn't create a
large impact on the board-specific code since the board only needs to
implement a single routine that returns a zero or non-zero value:
is_pci_host().

Justification for this becomes clear when considering systems running
in a cPCI environment:

1. Arbiter strapping: Many cPCI boards provide an external arbiter (often
part of the PCI-to-PCI bridge). Even though the arbiter is external (the
arbiter strapping is negated), the CPU may still be required to perform
local PCI bus configuration.

2. Host only: PPMC boards must sample the MONARCH# signal at run-time.
Depending on the configuration of the carrier boar, the PPMC board must
determine if it should configure the PCI bus at run-time. And in most
cases, access to the MONARCH# signal is board-specific (e.g. via
board-specific FPGA registers, etc).

In any event, the is_pci_host() routine gives each board the opportunity
to decide at run-time. If your board is always configured a certain way,
then just hardcode a return of 1 or 0 as appropriate.


Regards,
--Scott
<smcnutt@artesyncp.com>