Digital Equipment Corporation
Maynard, Massachusetts
September 1995
While Digital believes the information in this publication is correct as of the date of publication,
it is subject to change without notice.
Digital Equipment Corporation makes no representations that the use of its products in the
manner described in this publication will not infringe on existing or future patent rights, nor
do the descriptions contained in this publication imply granting of licenses to make, use, or sell
equipment or software in accordance with the description.
DEC, DECchip, Digital, Digital Semiconductor, OpenVMS, VAX DOCUMENT, and the
DIGITAL logo are trademarks of Digital Equipment Corporation.
Digital Semiconductor is a Digital Equipment Corporation business.
Intel is a trademark of Intel Corporation.
OSF/1 is a registered trademark of Open Software Foundation, Inc.
Prentice Hall is a registered trademark of Prentice-Hall, Inc. of Englewood Cliffs, NJ.
Windows NT is a trademark of Microsoft Corporation.
All other trademarks and registered trademarks are the property of their respective owners.
This document was prepared using VAX DOCUMENT Version 2.1.
•Chapter 8, DECchip 21171-BA Architecture Overview, provides an overview
of the DSW architecture.
xiii
•Appendix A, Technical and Ordering Information, provides information
on how to obtain technical support and how to order products and
documentation.
Document Conventions
This section describes the abbreviation and notation conventions used
throughout this guide.
Numbering
All numbers are decimal or hexadecimal unless otherwise specified. In cases
of ambiguity, a subscript indicates the radix of nondecimal numbers. For
example, 19 is decimal, but 1916and 19A are hexadecimal.
UNPREDICTABLE and UNDEFINED Definitions
Results specified as UNPREDICTABLE may vary from moment to moment,
implementation to implementation, and instruction to instruction within
implementations. Software can never depend on results specified as
UNPREDICTABLE.
Operations specified as UNDEFINED may vary from moment to moment,
implementation to implementation, and instruction to instruction within
implementations. The operation may vary from nothing to stopping system
operation. UNDEFINED operations must not cause the processor to hang, that
is, reach a state from which there is no transition to a normal state where the
machine can execute instructions.
xiv
Note the distinction between results and operations. Nonprivileged software
cannot invoke UNDEFINED operations.
Ranges
Ranges are specified by a pair of numbers separated by two periods (..) and are
inclusive. For example, a range of integers 0..4 includes the integers 0, 1, 2, 3,
and 4.
Extents
Extents are specified by a pair of numbers in angle brackets (<>) separated by
a ( :) and are inclusive. For example, bits <7:3> specifies an extent including
bits 7, 6, 5, 4, and 3.
Must Be Zero
Fields specified as must be zero (MBZ) must never be filled by software with a
non-zero value. If the processor encounters a non-zero value in a field specified
as MBZ, a reserved operand exception occurs.
Should Be Zero
Fields specified as should be zero (SBZ) should be filled by software with a zero
value. These fields may be used at some future time. Non-zero values in SBZ
fields produce UNPREDICTABLE results.
Register and Memory Figures
Register figures have bit and field position numbering starting at the right
(low-order) and increasing to the left (high-order). Memory figures have
addresses starting at the top and increasing toward the bottom.
Register Field Notation
Register figures and tabulated descriptions have a mnemonic that indicates the
bit or field characteristic as described in Table 1.
Table 1 Register Field Notation
NotationDescription
RWA read-write bit or field. The value can be read and written by software,
ROA read-only bit or field. The value can be read by software, microcode,
WOA write-only bit. The value can be written by software and microcode.
WZA write-only bit or field. The value can be written by software or
WCA write-to-clear bit. The value can be read by software or microcode.
RCA read-to-clear field. The bit is written by hardware and remains
microcode, or hardware.
or hardware. The bit is written by hardware. Software or microcode
write operations to this bit are ignored.
The bit is read by hardware. Read operations to this bit by software or
microcode return an UNPREDICTABLE result.
microcode. The bit is read by hardware. Read operations to this bit by
software or microcode return a zero.
Software or microcode write operations with a 1 to this bit cause the bit
to be cleared by hardware. Software or microcode write operations with
a 0 to this bit do not modify the state of the bit.
unchanged until the bit is read. The bit can be read by software or
microcode, at which point hardware can write a new value into the
field.
xv
Other register fields that are unnamed may be labeled as specified in Table 2.
Table 2 Unnamed Register Field Notation
NotationDescription
0A 0 in a bit position indicates a register bit that is read as a 0 and is
1A 1 in a bit position indicates a register bit that is read as a 1 and is
xAn x in a bit position indicates a register bit that does not exist in
ignored on a write operation.
ignored on a write operation.
hardware. The value is UNPREDICTABLE when read and is ignored
on a write operation.
Bit Notation
Multiple bit fields are shown as extents (see Extents).
Warning
Warning provides information to prevent personal injury.
Caution
Caution indicates potential damage to equipment or data.
Note
Note provides general information that could be helpful.
xvi
Data Units
Table 3 defines the data unit terminology used throughout this manual.
Table 3 Data Units
TermWordsBytesBitsOther
Byte1—8—
Word1216—
Tribyte—324—
Longword2432—
Quadword4864—
Octaword816128Single read fill; that is, the cache space
Hexword1632256Cache block, cache line. The space
that can be filled in a single read access.
It takes two read accesses to fill one
Bcache line (see Hexword).
allocated to a single cache block.
Signal Name References
Signal names in text are printed in boldface lowercase. Mixed-case and
uppercase signal naming conventions are ignored. These two examples
illustrate the conventions used in this document:
•MEM_WE_L[1] is shown as mem_we_l<1>.
•TEST_MODE[1] is shown as test_mode_h<1>.
xvii
DECchip 21171 Core Logic Chipset
The DECchip 21171 core logic chipset (21171) is used to support the Alpha
21164 microprocessor (21164) in high-performance uniprocessor systems. The
chipset includes an interface to the 64-bit peripheral component interconnect
(PCI) bus, and associated control and data paths for the 21164, memory, and
backup cache (Bcache). Little discrete logic is needed on the module.
1.1 Features
The DECchip 21171 core logic chipset has the following features.
One DECchip 21171-CA chip—The control, I/O interface, and address
chip (CIA) has 383 pins and is in a plastic pin grid array (PPGA)
package.
Four DECchip 21171-BA chips—The data switch chip (DSW) has 208
pins and is in a plastic quad flat pack (PQFP) package.
1
Overview
•Data paths
64-bit, ECC-protected data path between CIA and DSW
128-bit ECC-protected data path between the 21164 and DSW
256-bit ECC-protected memory data path between DSW and memory
•Timing
Supports all 21164 clock frequencies
Supports system clock frequencies up to 33 MHz (PCI clock)
PCI clock, cache, and memory clock period are an integer multiple of
the 21164 clock period
Preliminary Edition—Subject to Change—September 1995 1–1
DECchip 21171 Core Logic Chipset Overview
1.1 Features
•Bcache
Third-level cache for the 21164
Cache size from 2MB to 64MB
Write-back, ECC-protected cache
•Memory
Supports up to 8GB of ECC-protected physical memory
Industry-standard memory array configurations
•High-performance PCI
64-bit multiplexed address and data
64-bit PCI address handling
Scatter-gather map support
No glue logic required to connect PCI-compliant devices
1.2 System Overview
To build a system, such as that shown in Figure 1–1, with the DECchip 21171
core logic chipset, the following additional components are required:
•Alpha 21164 microprocessor
•Bcache
•Memory arrays
•Serial ROM (SROM) (for 21164)
•Circuit to generate input clock for 21164
•Circuit to generate input clock for 21171
•Support logic
PCI interrupt controller
PCI arbiter
Reset logic
Console ROM logic and flash ROM
1–2 Preliminary Edition—Subject to Change—September 1995
DECchip 21171 Core Logic Chipset Overview
1.2 System Overview
•PCI peripheral devices
Figure 1–1 DECchip 21171 Core Logic Chipset System Block Diagram
DECchip 21171
Bcache
Index
Alpha 21164
Microprocessor
SROM
sysBus
Data
sysBus Address
and Command
Core Logic Chipset
Four DECchip
21171-BAs
(Four DSWs)
IOD Bus
DECchip
21171-CA
(CIA)
MEMDATA Bus
Memory
Arrays
Memory Control
(ras, cas, and so on.)
PCI Bus
LJ-04214.AI
The MEMDATA bus provides a 256-bit, ECC-protected data path between the
four data path slices (DSWs) and the memory arrays. Each bit slice (DSW)
connects to 64 bits of the MEMDATA bus.
The input/output data (IOD) bus provides a 64-bit, ECC-protected data path
between the CIA and DSW chips.
The 21164 is connected to the CIA and DSW by the sysBus. The sysBus
contains address, data, and command signal lines. The sysBus is the 21164
interface bus with several additional signals to handle direct memory access
(DMA) transactions through the CIA and DSW.
Systems built with the DECchip 21171 core logic chipset have the capability
to support up to 8GB of memory. The chipset supports the following industry
standard dynamic RAMs (DRAM) in the memory arrays:
1MB3362MB336
4MB3368MB336
Preliminary Edition—Subject to Change—September 1995 1–3
As shown in Figure 1–1, the DECchip 21171 core logic chipset consists of the
CIA and four DSW application-specific integrated circuits (ASICs).
•DECchip 21171-CA (CIA), 383-pin PPGA: Provides control functions to
main memory, and a bridge to the 64-bit PCI bus. In addition, the CIA
provides all control functions for the DSW and part of the I/O data path.
•DECchip 21171-BA (DSW), 208-pin PQFP: A bit slice ASIC (four required),
provides the data path between the 21164, main memory, Bcache, and the
CIA, and part of the I/O data path.
1.3 CIA Microarchitecture
The CIA provides control functions to main memory and the PCI interface. As
shown in Figure 1–2, the CIA is divided into the following major functional
logic blocks.
The instruction and address logic accepts commands from the 21164 and
directs instructions to either the memory port or the I/O port. A 3-entry 21164
instruction queue provides buffering in the event that the memory and I/O
ports are busy. A flush address register is also provided to allow DMA read
and write operations to interrogate a supported Bcache for the most recent
data.
The memory logic provides the row and column addresses for the supported
memory banks and all memory control functions (row address select [ras],
column address select [cas], write-enable, and so on). In addition, the memory
logic provides control signals to the DSW to initiate data transactions to and
from memory.
The I/O address logic handles I/O read and write addresses. The decode
logic extracts the PCI address from the dense or sparse space 21164 address
encoding, generates PCI I/O read or write addresses, and passes them to the
PCI data path logic. The I/O logic can also increment the current address each
data cycle. The incremented address provides a pointer to the next data item
to be transferred to or from the I/O read/write buffers.
A 3-entry I/O address queue is provided together with a single-entry current
address register. These components allow four I/O write operations to be
outstanding. The DSW chip provides four corresponding I/O write data
buffers.
1–4 Preliminary Edition—Subject to Change—September 1995
DECchip 21171 Core Logic Chipset Overview
Figure 1–2 CIA Block Diagram
1.3 CIA Microarchitecture
Address
Path
Alpha 21164
Microprocessor
sysBus
Address/Command
Path
SROM
DECchip 21171-CA (CIA)
DMA Read/Write and
Instruction and
Address Logic
I/O Port
Memory
Logic
Memory
Control
Functions
Memory ArraysIOD Bus
TLB Miss Address
I/O Address
Logic
PCI I/O
Read/Write
Address
DMA Address
PCI Data
Path Logic
(To/From DSW)
Logic
PCI
Bus
LJ-04215.AI
In addition, a bypass path is provided for the special case of a dense space I/O
write transaction with all 32 data bytes valid.
The PCI data path logic provides the interface to the 64-bit PCI, and contains
a portion of the data path control logic. Its major functions are to provide:
•ECC code generation on data transactions to the DSW
•ECC code check and correction on data transactions from the DSW
•I/O read/write addresses to the PCI
•Buffers for all I/O and DMA read and write data transactions
Preliminary Edition—Subject to Change—September 1995 1–5
DECchip 21171 Core Logic Chipset Overview
1.3 CIA Microarchitecture
The DMA address logic converts PCI addresses to 21164 memory address
space. Two address conversion methods are provided:
•A direct path where a base offset is concatenated with the PCI address
•A scatter-gather map, which maps any 8KB PCI page to any 8KB memory
space page
The scatter-gather map translation lookaside buffer (TLB) has eight entries,
with each entry containing four consecutive page table entries (PTEs).
1.4 DSW Microarchitecture
The DSW provides bidirectional data paths between the 21164, the memory
arrays, and the CIA, as shown in Figure 1–3. Four DSW chips, using a bit
slice design, provide the complete data path.
The majority of the DSW logic is comprised of data buffers and multiplexers.
The CIA, using two encoded control fields, directs data flow to and from the
DSW.
The victim buffer has a 64-byte single-entry buffer to contain victim data
during a 21164 memory read transaction that generates a victim. Victim data
is saved in the buffer, allowing the read transaction to fetch read data from
memory and send it to the 21164. Later the victim data is written to memory.
The memory interface contains the data-out register and the data-in register.
The data-out register buffers either 21164 victim data to memory, or DMA
write data to memory. The data-in register accepts 21164 memory read data.
The DSW clocks the register on optimum 15-ns clocks to minimize memory
latency.
The I/O write buffer consists of four 32-byte buffers. The four buffers allow
the chipset to sustain maximum bandwidth on a large copy operation from
memory, through the 21164, to I/O space.
The 32-byte I/O read buffer is provided to assemble the 128-bit data required
by the 21164 from the 64-bit data transfers from the IOD bus (CIA-DSW
interface).
The DMA read/write buffers consist of two identical buffers (DMA buffer 0 and
DMA buffer 1). Each buffer consists of three, 64-byte, single-entry data buffers
used to hold data from memory, Bcache, and the PCI. When a buffer is written,
its data valid bits are asserted, along with data, to the buffer. The valid bits
are used to select appropriate read/write data.
1–6 Preliminary Edition—Subject to Change—September 1995
Figure 1–3 DSW Block Diagram
DECchip 21171 Core Logic Chipset Overview
1.4 DSW Microarchitecture
Alpha 21164
Microprocessor
SROM
Four DECchip 21171-BAs (Four DSWs)
Victim Buffer
21164 Read Data
21164
Interface
I/O
Write Buffer
I/O Write
Data Path
I/O
Read Buffer
IOD Bus Interface
IOD Bus (To/From CIA)
Memory
Arrays
Memory
Interface
DMA
Buffer
LJ-04216.AI
1.5 System Clocks
The 21164 internal clock will be divided by a programmable value to create an
output clock, sys_clk_out1_h, which is used as the system clock. The divisor
must produce a clock between 25 MHz and 33 MHz.
The system designer must supply a circuit to maintain stability of the system
clock and distribute it to PCI devices.
The system clock can be used as a source for a circuit that generates a clock
for (E)ISA devices.
Preliminary Edition—Subject to Change—September 1995 1–7
Power supplyVss 0.0 V, Vdd 5 V ±5%
Operating temperatureTj maximum = 70°C (154°F)
Storage temperature range–55°C (–67°F) to 125°C (257°F)
Power dissipation
Power supplyVss 0.0 V, Vdd 5 V ±5%
Operating temperatureTj maximum = 100°C (212°F)
Storage temperature range–55°C (–67°F) to 125°C (257°F)
Power dissipation
@Vdd = 5.25 V, Freq = 33 MHz
1–8 Preliminary Edition—Subject to Change—September 1995
1.0 W maximum
PartI
DECchip 21171-CA (CIA) Information
Part I contains five chapters that provide information about the DECchip
21171-CA (CIA) chip. The following table lists the topics described in each
chapter:
ChapterDescription
2Pin signals
3Architecture
4Control and status registers
5Power-up and initialization
6System address mapping
DECchip 21171-CA Pin Descriptions
This chapter provides a description of the DECchip 21171-CA pin signals.
2.1 DECchip 21171-CA Pin List
Table 2–1 lists the pin signals grouped by function. The information in the
Type column identifies a signal as input (I), output (O), bidirectional (B), or
power (P).
2–2 Preliminary Edition—Subject to Change—September 1995
(continued on next page)
DECchip 21171-CA Pin Descriptions
2.1 DECchip 21171-CA Pin List
Table 2–1 (Cont.) DECchip 21171-CA Pin List
Signal NameQuantityTypeSignal NameQuantityType
Reserved Signals (2 Total)
gru_ack_h1Igru_sel_h1O
Pin Totals
Total input pins:
Total output pins:
Total bidirectional pins:
Total signal pins:
21
72
196
—289
Total signal pins:
Total PWR5 pins:
Total GND pins:
Total miscellaneous power pins:
Total spare pins:
Total reserved pins:
Total pins:
289
36
51
4
1
2
—383
2.2 DECchip 21171-CA Signal Descriptions
This section provides a description of the pin function. Signal descriptions are
grouped by function and correspond to the pin list (see Table 2–1).
Note
The Alpha 21164 microprocessor uses the rising edge of
sys_clk_out1_h (clk1R) as a reference when generating and sampling
the system interface signals.
2.2.1 sysBus Signals
The Alpha 21164 Microprocessor Hardware Reference Manual defines and
describes sysBus signals in detail, except for int_h and error_h. Those two
signals are described here.
Preliminary Edition—Subject to Change—September 1995 2–3
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
int_h
The int_h pin is connected to 21164 pin irq<0>. The CIA asserts int_h to
notify the 21164 that the CIA has detected a corrected ECC error.
error_h
The error_h pin is connected to 21164 pin irq<3>. The CIA asserts this
signal line to notify the 21164 that the CIA has detected an error other than a
corrected ECC error.
2.2.2 PCI Local Bus Signals
The PCI local bus signals
ack64_l
Acknowledge 64-bit transfer—Asserted by the device that has decoded its
address as the target of the current access, indicating that the target device is
willing to transfer 64-bit data.
ad_h<63:0>
The 64-bit address and 64 bits of data are multiplexed on these signal lines.
The upper 32 bits (<63:32>) are reserved except for those transfers where
req_64_l is asserted.
cbe_l<7:0>
Bus command and byte enable are multiplexed on the same signal lines.
During an address phase (when using the DAC command and when req_64_l
is aserted) the bus command is transferred on cbe_l<7:0>. During a data
phase these lines carry the byte enable bits, indicating which bytes have
meaningful data.
devsel_l
When asserted, indicates that the asserting device has decoded its address as
the target of the current access. As an input, devsel_l indicates whether any
device on the bus has been selected.
frame_l
The cycle frame signal is asserted by the current master to indicate the
beginning and duration of an access. When frame_l is deasserted, the final
phase of the access has started.
gnt_l
Assertion of gnt_l indicates that access to the bus has been granted. This is a
point-to-point signal. Every master has its own grant line.
2–4 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
irdy_l
Assertion of irdy_l (initiator ready) indicates that the bus master is ready
to complete the current phase of the transaction. Signal irdy_l is used in
conjunction with trdy_l. A data phase is completed on any clock where both
signals are sampled asserted.
lock_l
Assertion of lock_l indicates an atomic operation that may require multiple
transactions to complete. When lock_l is asserted, nonexclusive transactions
may proceed to an address that is not currently locked.
mem_cs_l
The PCI Local Bus Specification does not preclude product-specific function
/performance enhancements by way of sideband signals. A sideband signal
is loosely defined as any signal that is not part of the PCI specification
and connects two or more PCI nodes and has meaning only to these nodes.
Sideband signals lines must not violate PCI local bus specifications.
Signal mem_cs_l is a sideband signal line. It is used by the PCI-to-EISA
bridge to help manage the EISA memory map and peripheral component
architecture (PCA) compatibility holes (see Section 6.3.4.2).
par_h
The par_h signal is used to create even parity across ad_h<31:0> and
cbe_l<3:0>. Signal par_h is stable and valid one clock cycle after the address
phase. Signal par_h remains valid until one clock cycle after the completion of
the current data phase. The master asserts par_h for address and write data
phases; the target asserts par_h for read data phases.
par64_h
The par64_h signal is used to create even parity across ad_h<63:32> and
cbe_l<7:4>. Signal par64_h is stable and valid one clock cycle after the
address phase. Signal par64_h remains valid until one clock cycle after the
completion of the current data phase. The master asserts par64_h for address
and write data phases; the target asserts par64_h for read data phases.
perr_l
The perr_l signal (parity error) is used to report data parity errors that
occur during all PCI transactions except a special cycle. Signal perr_l is a
tristate signal line that must be asserted by the data-receiving node two cycles
after data with bad parity is detected. Signal perr_l must be asserted for a
minimum of one clock cycle for each phase during which bad data parity is
detected. Signal perr_l, like all sustained tristate signals, must be asserted
high for at least one cycle before being returned from asserted to the tristate
condition.
Preliminary Edition—Subject to Change—September 1995 2–5
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
req_l
Assertion of the req_l signal line (request) indicates to the arbiter that this
node wants to use the PCI bus. Every master node has its own req_l pin.
req64_l
The current bus master asserts req64_l (request 64-bit transfer), indicating
that it wants to transfer data by using 64 bits. Signal req64_l has the same
timing as frame_l. Signal req64_l is asserted low by the system reset signal.
Nodes connected to the 64-bit PCI bus extension will see the line asserted low.
Those nodes not connected to the 64-bit PCI bus extension will see the signal
asserted high by a pull-up device.
rst_l
The rst_l signal line (reset) is used to bring PCI-specific registers, sequencers,
and signals to a consistent state. When rst_l is asserted, all PCI output pins
must be driven to their benign state, usually tristated. Although the signal is
asynchronous, deassertion must be a clean, bounce-free edge.
serr_l
The serr_l pin (system error) is used to report address parity errors, data
parity errors on the special cycles command, or any other system error where
the result will be catastrophic. If a node does not want a nonexistent memory
(NMI) error to be generated, a different reporting mechanism is required.
serr_l is an open-drain signal line and is asserted for a minimum of one clock
cycle by the node reporting the error.
stop_l
The stop_l pin indicates that the current target is requesting that the master
stop the current transaction.
trdy_l
When asserted, the trdy_l (target ready) pin indicates the target node’s
ability to complete the current phase of the transaction. Signal trdy_l is
used in conjunction with irdy_l. A data phase is completed on any clock
assertion where both trdy_l and irdy_l are sampled asserted. During a read
transaction, trdy_l indicates that valid data is present on ad<31:0>. During
a write transaction, trdy_l indicates that the target is ready to accept data.
Wait cycles are inserted until both trdy_l and irdy_l are asserted together.
2–6 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
2.2.3 IOD Bus Signals
This section provides a description of the IOD bus signals.
iod_h<63:0>
The iod_h<63:0> signals carry data between the CIA and DSW on the IOD
bus.
iod_e_h<7:0>
The iod_e_h<7:0> signals carry ECC between the CIA and DSW on the IOD
bus. All ECC generation and detection is performed in the 21164 and CIA.
ioc_h<6:0>
The ioc_h<6:0> signals allow the CIA to control data flow on the IOD bus. The
data transfers are between memory and the PCI bus. CIA control is encoded
as shown in Table 2–2.
Table 2–2 ioc_h<6:0> Control Functions
ioc_h
<6:4>
000Buffer nxxxClear valid bits in PCI buffer n.
001xxxxLoad the LW register (Reserved).
010Buffer nAdr m2Adr m1Adr m0Write PCI buffer n, adr m.
011Mask3Mask2Mask1Mask0Write IOR (QW mask).
100Buffer nAdr m2Adr m1Adr m0Read DMA buffer n, adr m.
101Buffer n1Buffer n0Adr m1Adr m0Read IOW buffer n, adr m.
110Buffer n1
111Mask3Mask2Mask1Mask0Start IOW (buffer mask).
ioc_h
<3>
Buffer n
ioc_h
<2>
Buffer n0
Adr m2
ioc_h
<1>
Adr m1Adr m0Read buffer (IOW or DMA) low
ioc_h
<0>Function
longword.
Preliminary Edition—Subject to Change—September 1995 2–7
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
cmc_h<8:0>
The cmc_h<8:0> signal lines control flow of data between the 21164 and the
memory arrays. They also start and stop loading of the free-running buffers
(victim, flush, IOW buffers) in the DSW. In Table 2–3, A5 and A4 are address
bits. The Var field encodes the transaction being performed as shown in
Table 2–4, Table 2–5, and Table 2–6.
Table 2–3 cmc_h<8:0> Control Functions
cmc_h
<8:5>
0000Var 4
0010xA4
0011xA4xVar 1Var 0Move IOR buffer data to
0100A5
0101A5A4DelayVar 1Var 0Move memory data to
0110A5A4DelayStop
0111A5A4DelayStop
1000A5A4Var 2Var 1Var 0Write victim buffer to
1010A5A4Var 2Var 1Var 0Write DMA0 buffer to
1011A5A4Var 2Var 1Var 0Write DMA1 buffer to
cmc_h
<4>
2
cmc_h
<3>
1
Var 3Var 2Var 1Var 0Start/stop the buffer.
2
cmc_h
<2>
cmc_h
<1>
cmc_h
<0>Function
DelayVar 1Var 0Move fill data from
memory to 21164.
21164.
A4DelayVar 1Var 0Move memory data to
memory buffer 0.
memory buffer 1.
IOWn1
Stop
IOWn0
Move memory data to
memory buffer 0, stop
IOW buffer n.
IOWn1
Stop
IOWn0
Move memory data to
memory buffer 1, stop
IOW buffer n.
memory.
memory.
memory.
1
The variable field encodes the transaction being performed. See Table 2–4, Table 2–5, and Table 2–6.
2
A5 and A4 are address bits.
2–8 Preliminary Edition—Subject to Change—September 1995
Preliminary Edition—Subject to Change—September 1995 2–9
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
2.2.4 Memory Control Signals
This section provides a description of the memory control signals.
cas_h<3:0>
The cas_h<3:0> signal lines select the column address in the memory arrays.
mem_ack_l
The mem_ack_l PCI sideband signal line is asserted by the CIA when it has
cleared the transaction path in preparation to receive a transaction from the
PCI-to-EISA bridge. See mem_req_l.
mem_addr_h<12:0>
The mem_addr_h<12:0> signal lines carry the memory array address.
mem_en_h
The mem_en_h signal line enables operation of the memory arrays.
mem_req_l
The mem_req_l signal line is asserted by the PCI-to-EISA bridge to notify
the CIA that it has a transaction pending. The CIA completes all transactions
in progress and then asserts mem_ack_l to the PCI-to-EISA bridge. The
PCI-to-EISA bridge then starts the transaction. See mem_ack_l.
The PCI-to-EISA bridge set can be made using these two chips:
•Intel 82374EB EISA System Component (ESC)
•Intel 82375EB PCI-to-EISA Bridge (PCEB)
The PCEB allows a certain time period for a PCI transaction to complete. The
mem_req_l and mem_ack_l PCI sideband signal lines are used to ensure that
that the CIA performs transactions from the PCI-to-EISA bridge immediately.
mem_we_l<1:0>
The mem_we_l<1:0> signal lines carry the write-enable signal to the memory
arrays.
ras_h<3:0>
The ras_h<3:0> signal lines select the row address in the memory arrays.
set_sel_h<15:0>
The set_sel_h<15:0> signal lines are used to select a memory array.
2–10 Preliminary Edition—Subject to Change—September 1995
2.2.5 Phase-Locked Loop Signals
This section describes the phase-locked loop signals.
pll_agnd_h
Signal pll_agnd_h is the analog ground input to the PLL and is connected to
the system board ground.
pll_clk_h
The pll_clk_h signal is the system clock (or reference signal) input to the
phase-locked loop (PLL) circuit. The pll_clk_h signal line is the source for CIA
timing. It supplies a clock with the same frequency as that supplied to the
PCI.
pll_lp1_h
Signal pll_lp1_h is an output of the PLL (phase detector output).
pll_test_h<3:0>
Signals pll_test_h<3:0> select which internal chip information is presented at
output pins debug_h<1:0>. These signal lines are reserved to Digital.
pll_vss_h
Signal pll_vss_h is the ground input input to the PLL. It is connected to the
system board ground.
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
2.2.6 Miscellaneous Power Signals
This section describes the miscellaneous power signals.
aux_vdd_h
Signal aux_vdd_h is modified (filtered) within the CIA chip before going to the
PLL section of the CIA. Signal aux_vdd_h should be connected to vdd_h on
the module.
aux_vss_h
Signal aux_vss_h is modified (filtered) within the CIA chip before going to the
PLL section of the CIA. Signal aux_vss_h should be connected to vss_h on the
module.
pll_vdd_h
Signal pll_vdd_h is the power input to the PLL. It is connected to the system
board power.
pll_lp2_h
Signal pll_lp2_h is an input to the PLL, driving a VCO input.
Preliminary Edition—Subject to Change—September 1995 2–11
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CA Signal Descriptions
2.2.7 Miscellaneous Signals
This section provides a description of the miscellaneous signals.
count_out_h
Signal count_out_h is used during PLL tests.
debug_h<1:0>
Signals debug_h<1:0> present internal CIA information selected by
pll_test_h<3:0>. These signal lines are reserved to Digital.
scan_in_h
Signal scan_in_h is the enable pin for the scan chain.
sys_rst_l
The sys_rst_l signal line is the system reset input signal to the CIA. When
sys_rst_l is asserted, the CIA resets its internal state.
test_mode_h<1:0>
Signals test_mode_h<1:0> define the four operating modes of the CIA as
listed in Table 2–7.
Table 2–7 DSW Operating Modes
test_mode_h<1:0>DSW Operating Mode
00Tristate mode
01Normal mode
10Scan mode
11PLL test mode
test_or_scan_out_h
Signal test_or_scan_out_h is the TEST_OUT output that is the end of the
parametric NAND tree.
2.2.8 Reserved Signals
This section provides a description of the reserved signals.
gru_ack_h
Signal gru_ack_h is reserved for use by Digital.
gru_sel_h
Signal gru_sel_h is reserved for use by Digital.
2–12 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Pin Descriptions
2.2 DECchip 21171-CASignal Descriptions
2.2.9 DECchip 21171-CA Pin Name List (Alphabetic)
Table 2–8 lists the DECchip 21171-CA (CIA) pins by name in alphabetical
order. The following abbreviations are used in the Type column of the table:
Figure 2–1 and Table 2–10 indicate the DECchip 21171-CA package
dimensions.
2–24 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Pin Descriptions
2.3 DECchip 21171-CA Mechanical Specifications
Figure 2–1 DECchip 21171-CA Package Dimensions
D
M
D2
E
(Note 3)
-C-
(Note 3)
H
E2
A2
(Note 3)
A1
(Note 2)
M1
AB
AA
Y
W
V
U
T
R
P
N
M
N
L
K
J
H
N1
G
F
E
D
C
B
A
2345678910111213141516171819202122
4X Stand-Off Pin
J
(Note 1)
L1
Q1
D3
D1
(Note 2)
G
E1
(Note 2)E3(Note 4)
A
L
B1
B
Notes for Figure 2–1
The following notes apply to Figure 2–1:
Note 1. Chip capacitors envelope. Commercial grade chip caps; nominal
capacitance 0.082F.
Note 2. Envelope for maximum lid (ceramic, epoxy seal).
Note 3. Envelope for integral heat slug.
Note 4. The drawing/dimensions are for reference only. For board layout,
request an engineering drawing from Digital Semiconductor.
Preliminary Edition—Subject to Change—September 1995 2–25
F
Seating
Plane
LJ-04217.AI
DECchip 21171-CA Pin Descriptions
2.3 DECchip 21171-CA Mechanical Specifications
Table 2–10 DECchip 21171-CA Package Dimensions
IndicatorMinimumMaximum
A1.78 mm (0.070 in)3.30mm (0.130 in)
A1NA
A2NA1.27 mm (0.050 in)
B0.40 mm (0.016 in)0.51 mm (0.020 in)
B11.17 mm (0.046 in)1.37 mm (0.054 in)
D57.30 mm (2.256 in)57.51 mm (2.264 in)
D1NA24.89 mm (0.980 in)
D2NA24.13 mm (0.950 in)
E57.30 mm (2.256 in)57.51 mm (2.264 in)
E1NA24.89 mm (0.980 in)
E2NA24.13 mm (0.950 in)
G1.78 mm (0.070 in)2.04 mm (0.080 in)
HNA3.56 mm (0.140 in)
INA4.06 mm (0.160 in)
JNA1.65 mm (0.064 in)
L4.06 mm (0.160 in)4.57 mm (0.180 in)
L11.12 mm (0.044 in)1.42 mm (0.056 in)
Q10.23 mm (0.009 in)NA
1
4.19 mm (0.165 in)
IndicatorBasic
D32.100 in (53.34 mm)
E32.100 in (53.34 mm)
F1.0 in (2.54 mm)
M2.050 in (52.07 mm)
M11.000 in (25.4 mm)
N2.000 in (50.8 mm)
N10.950 in (24.13 mm)
1
NA = Not applicable.
2–26 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
This chapter describes the DECchip 21171-CA (CIA) data paths with functional
units, transactions, and architecture.
3.1 CIA Data Paths with Functional Units Description
This section shows the CIA data path functional units and describes operation
of the functional units.
3.1.1 System Functional Units
The CIA operates as the host bridge on a PCI-based system along with the
21164, four DSWs, Bcache, memory arrays, and support logic. A simplified
description of the system data paths and functional units is presented here to
show where the CIA operates within the system.
The CIA performs control, I/O, and address tasks within the DECchip 21171
Core Logic Chipset as shown in Figure 3–1. The host bridge module is shown
plugged into a 64-bit PCI.
3
Preliminary Edition—Subject to Change—September 1995 3–1
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
Figure 3–1 System Functional Block Diagram
Bcache
21164
PCI-to-(E)ISA
Bridge
(E)ISA
Bus
Data
Switch
CIA ASIC
32-Bit
Slots
Memory
64-Bit PCI Bus
PCI-to-PCI
Bridge
Slot
Slot
Slot
Audio
64-Bit Slots
PCI
Graphics
Internal PCI
SCSISCSISCSI
Ethernet
3–2 Preliminary Edition—Subject to Change—September 1995
LJ-04218.AI
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
3.1.2 CIA Functional Units
Figure 3–2 shows the CIA data paths and functional units.
Figure 3–2 CIA Functional Block Diagram
21164 Address
and Command
(IOD Bus)
21164
Instruction
and
Address
Logic
Mem PortI/O Port
21164
Instruction
Queue
Memory
Logic
Memory
RAS
Check
Control
Data_Out
Flush
Address
DMA Read/Write and TLB Miss Address
Out
MUX
ECC Gen
I/O Address Logic
Bypass Path
I/O Address
Queue
Data_In
DMA Write/
Merge Path
I/O Read
Path
DMA Write
Path
DMA Write
Address
Decode
and
Increment
ECC Corr
I/O Read
DMA Read
I/O Write
I/O Write
PCI Data Bypass
I/O Read/Write
Address for PCI
DMA Address
Logic
PA Reg.
Big MUX
PCI Data
PCI
Data-Path
Logic
PCI Address
PCI
Target
Window
Logic
PCI Data/
Address Reg.
Scatter-Gather TLB
Direct Map
TLB Miss Address
PCI Bus
Address/
Command
Hit
Memory Address,
ras, we, Refresh
Current
Address
Register
DMA Read
Prefetch Counter
LJ-04219.AI
Preliminary Edition—Subject to Change—September 1995 3–3
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
The CIA functional units are listed here and are described in Section 3.1.2.1
through Section 3.1.2.5:
•21164 instruction and address logic
•PCI data path logic
•Memory logic
•I/O address logic
•DMA address logic
3.1.2.1 21164 Instruction and Address Logic
The 21164 instruction and address logic accepts commands from the 21164 and
directs the instruction to the memory port or the I/O port. The DMA read/write
address (or the scatter-gather TLB miss servicing address) also has access to
the memory port through a multiplexer.
The 21164 instruction queue is provided to capture 21164 commands should
the memory or I/O port be busy. The 21164 can issue, at most, two read
miss transactions—with one having a victim. This maximum condition would
require holding three addresses and commands, so a 3-entry buffer is needed.
The flush address register is used during DMA read or write transactions
to the Bcache for the latest data. The Bcache is implemented as a write
invalidate cache and so might have the only valid copy of the required data.
3.1.2.2 PCI Data Path Logic
Because the CIA is the PCI host bridge it generates and decodes PCI addresses
and also supplies and receives data. The data path to and from the PCI goes
through the CIA and the DSW. The ECC generation and check logic is excluded
from the DSW because of the bit slice nature of the DSW and so is performed
by the CIA.
Separate buffers are provided in the CIA for DMA and I/O read/write
transactions. These buffers are either 32 bytes or 64 bytes in size, and are
shown partitioned into 16-byte units, which matches the width of the 21164
data bus.
Later, when the DSW is examined, it will be seen that the DSW also has
buffers for the DMA and I/O paths. Control logic simplification is the reason
for this duplication.
3–4 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
Table 3–1 lists the PCI commands that the CIA responds to and sends.
•Issues PCI fast back-to-back cycles in dense address space
Preliminary Edition—Subject to Change—September 1995 3–5
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
3.1.2.3 Memory Logic
Memory logic provides the row and column address for the memory banks
and all the control signals (ras_h, cas_h, mem_we_l<1:0>, and mem_en_h).
The memory logic provides control signals to the DSW to inform it when to
send/strobe the data to/from memory. The CIA also controls memory refresh.
3.1.2.4 I/O Address Logic
The I/O address logic handles the I/O read and write addresses. The address
decode logic extracts the PCI address from the dense/sparse space 21164
address encodings (see Chapter 6 for more details).
This logic also increments the current address of each data cycle for two
reasons:
•In case of a PCI retry, which will require the transaction to be resumed
later, we need to start with the address of the aborted data.
•We need to provide a pointer to the next data item to be loaded/sent from
the I/O read/write buffers.
The CIA can queue up to six I/O write transactions. Two of the I/O write
transactions can be waiting in the 21164 queue and four I/O write transactions
can be sitting in the I/O address queue (a 3-entry I/O address queue is provided
together with a single-entry current address register). This allows six I/O write
transactions to be outstanding (the DSW provides four corresponding I/O write
data buffers). The bypass path is enabled/disabled by CIA_CTRL[CSR_IOA_
BYPASS] (CSR_CTRL<15>) being equal to 0/1 respectively.
3–6 Preliminary Edition—Subject to Change—September 1995
The CIA provides two more buffers that are needed to sustain maximum
bandwidth for memory copy to I/O space. A bypass path is provided for certain
dense-space I/O write transactions, that is, the valid bits for the first 16 bytes
of data are either 1111, 1011, 1010, and 1001—these cases are optimized for a
64-bit PCI with at least two PCI data cycles.
3.1.2.5 DMA Address Logic
The DMA address logic converts PCI addresses to CPU memory addresses.
Two conversion methods are provided:
•A direct path where a base offset is concatenated with the PCI address
•A scatter-gather map, which maps any 8KB PCI page to any 8KB memory
space page
The scatter-gather TLB is eight entries deep with each entry holding four
consecutive PTEs. A TLB miss is handled by hardware but software is required
to invalidate stale entries by writing to the translation buffer invalidate
register (TBIA). See Chapter 6 for more information on address mapping.
The output of the physical address register (PA) has a a counter that generates
the prefetch address for certain DMA read miss transactions. An 8KB detector
prevents prefetching across page boundaries.
DECchip 21171-CA Architecture Overview
3.1 CIA Data Paths with Functional Units Description
3.2 CIA Transactions
The CIA performs the transactions listed here:
•21164 memory read miss
•21164 memory read miss with victim
•21164 I/O read
•21164 I/O write
•DMA read
•DMA read (prefetch)
•DMA write
These transactions are described in detail in Section 3.2.1 through
Section 3.2.8.
Preliminary Edition—Subject to Change—September 1995 3–7
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
3.2.1 21164 Read Miss Transaction
The 21164 read miss command and address are sent to the CIA. If the CIA
is idle, the command will be stored directly in the memory port register;
otherwise, the command enters the 3-deep CPU instruction queue and remains
there until the memory port is free. The CIA can accept another subsequent
read miss command from the 21164 and store it in the 21164 instruction queue.
The three data paths to the memory port have to arbitrate for use of the
memory port: (1) the direct path and (2) the 21164 instruction queue previously
mentioned and (3) the DMA read/write address path. The DMA read/write
address path is also the path for scatter-gather TLB miss addresses. Upstream
of the multiplexer all instruction ordering from the 21164 is preserved.
Downstream of the multiplexer ordering between I/O write transactions and
21164 memory transactions is lost. This ‘‘post and run’’ I/O write coherency
issue is discussed in Section 3.3.2.
When the 21164 read miss command enters the memory logic, the memory
controller generates ras_h, cas_h, and other address signals for the memory
array. The memory controller then waits for the memory access delay before
instructing the DSW to accept the memory data. The MEMDATA bus width
is 256 bits (32 bytes), and so two memory cycles are required to access the
64-byte block.
The DSW synchronizes the data to the 21164 clock. The fixed 64-byte block
size read data is returned to the 21164 in wrapped-order.
3.2.2 21164 Read Miss with Victim Transaction
The read miss with victim transaction is similar to the read miss transaction
described in Section 3.2.1, except the 21164 will also use a Bcache victim
command to send out the victim block. The victim block is the modified (dirty)
block that is to be replaced in the Bcache by the read miss transaction data.
The read miss command is followed by the Bcache victim command and
address. The read miss command will get to the memory port first and so
the Bcache victim command and address are always written into the 21164
instruction queue. The victim block data is saved in the victim buffer in the
DSW.
3–8 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
The CIA arbiter ensures that a read miss transaction and a victim write
transaction are an atomic operation. The victim data is written to memory
after the read data has been fetched from memory. The row address portion
of the victim address is compared to the row address of the current read miss
transaction. If they match, then no memory ras_h cycle is required; instead,
only a cas_h cycle is needed. The address bits for the memory have been
carefully interspersed to maximize the chance of a victim row address ‘‘hitting’’
the read address—this performance feature is transparent to software. The
memory controller in the CIA generates the memory write pulses and instructs
the DSW to send the victim data to memory.
The DSW victim buffer is invalidated if a DMA write transaction (or DMA
read with lock transaction) ‘‘hits’’ in the victim buffer. Until the 21164 has its
read miss transaction data returned, the victim block is still in the Bcache
and is still ‘‘owned’’ by the 21164. It is possible for the read miss with victim
transaction from the 21164 to be stalled behind a DMA write transaction. The
DMA write transaction could ‘‘hit’’ the victim block—in this case, the DMA
write transaction will cause a flush command to be issued to the 21164, which
will result in the 21164 providing the victim data to the DSW and then the
21164 will invalidate the victim block in Bcache. Thus, the victim data waiting
in the victim buffer is no longer valid and is marked invalid by logic in the
CIA.
3.2.3 21164 Read Transaction to Noncacheable Space
A read transaction by the 21164 to noncacheable space can be to one of four
address spaces:
•PCI memory space
•PCI I/O dense or sparse space
•PCI configuration space
•CIA control and status register (CSR) space
Preliminary Edition—Subject to Change—September 1995 3–9
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
The read command to noncacheable space is accepted by the CIA like the read
miss command, except the instruction goes to the I/O port. All read commands
to noncacheable space go to the I/O address queue. There is no bypass path for
dense space I/O read transactions.
Decode logic is provided on the output of the I/O address queue to extract the
byte address for the PCI from the 21164’s sparse space encoding, and to decode
the address for the correct region (PCI memory, I/O, configuration, or CSR).
A read command to noncacheable space destined for the PCI is described here
because it is the more interesting case. Logic increments the current longword
/quadword address stored in the current address register each data cycle. This
is needed in case of a PCI retry and is also used to index the next data item
in the I/O write/read data buffers. The value in the current address register
corresponds to the address of the data item on the PCI bus (it has the same
canonical time as the PCI data-out register).
The I/O read command address is sent to the PCI after any prior I/O write
transactions have completed (strict ordering is maintained). The PCI returns
the requested data and places it in the I/O read buffer. The contents of this I/O
read buffer are next copied to the I/O read buffer in the DSW and then sent to
the 21164.
Why have the DSW I/O read buffers been replicated in the CIA? The I/O read
buffers in the CIA are provided for the following reasons:
•The path to the DSW might be busy with the end of a prior DMA write
transaction (especially if the data has to be merged with memory data to
build it up to the ECC width).
•At least 64 bits of storage are required for ECC because the PCI can return
32 bits of data at a time.
•It is convenient to handle transaction retries forced by the PCI protocol.
The I/O read buffer in the DSW was provided to build the data up from
64 bits (IOD bus data lines between the CIA and DSW) to the 128 bits
required by the 21164.
I/O read transactions from the 21164 have 8-byte resolution and the CIA
always returns 32 bytes. If smaller resolution is required, sparse space
addressing must be used. Sparse space addressing is described in Chapter 6.
Data is returned in the appropriate byte lanes.
3–10 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
3.2.4 21164 Write Transaction to Noncacheable Space
The 21164 issues write transactions to noncacheable space with longword
resolution. For smaller granularity, sparse space addressing must be used (see
Chapter 6).
Data for I/O write transactions is captured in the I/O write buffer in the DSW.
Four 32-byte buffers are available in the DSW and a further two in the CIA.
This number of entries allows the CIA to sustain maximum bandwidth on a
large copy operation from memory through the 21164-to-I/O space.
The data from these I/O write buffers is sent to the two I/O write buffers in the
CIA: two buffers are provided, allowing one to be emptied to the PCI bus while
the other is being filled. Each 32-byte buffer in the CIA requires a separate
PCI transaction (no merging of the write buffers occurs).
Another benefit of the CIA’s I/O write buffers is simpler data-flow management
when a target PCI device stalls the I/O write transactions.
The address for an I/O write transaction is sent to the CIA I/O port. For a
32-byte aligned dense space write transaction, the I/O write command is either
sent directly to the PCI bus via the bypass path, or queued up in the I/O
address queue. The direct (bypass) path is used if:
•There are no outstanding I/O commands.
•The command is an I/O write command to dense space and the first
16 bytes are mostly valid (like 1111, 1011, 1010, 1001).
This command also goes into the I/O address queue, but only as a convenient
path in the I/O addressing logic section, to get the address into the current
address register.
A total of six I/O write addresses can be queued: the first in the current
address register, two in the 21164 queue, and three in the I/O address queue.
The I/O address queue maintains strict ordering for I/O transactions but does
not maintain strict ordering of I/O write transactions relative to any memory
read or write transactions (see Section 3.3).
Preliminary Edition—Subject to Change—September 1995 3–11
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
3.2.5 DMA Transactions
The PCI address and command are captured in the address/command register
and the data/address register. The address is compared against four address
windows to determine if this PCI command should be accepted or ignored by
the CIA.
Address windows are a requirement of the PCI specification and are software
programmable—they are described in detail in Chapter 6. All PCI commands
destined for memory are accepted by the CIA.
Two registers capture incoming samples from the PCI bus.
•The address register captures the address and command, holding
them for the duration of the transaction.
•The data-in register captures the data, but also captures the
address.
The target window logic is attached to the data-in register. It is located
there to capture a buffer of data (64 bytes) for use with a DMA write
transaction scatter-gather TLB ‘‘miss.’’ The target window logic retries
the master PCI device in case it has more data.
The CIA has released the PCI for further transactions but is busy
servicing the TLB ‘‘miss’’ so must hold the virtual address in the
address register. Should another PCI transaction occur while the CIA
is servicing the TLB ‘‘miss,’’ the CIA will accept that address and
compare it to the target window.
Note
There are three registers associated with each of the four PCI windows:
•Window base register 0–3 (Wn_BASE)—defines the start of the target
window. Wn_BASE[Wn_BASE_SG] determines if the scatter-gather map is
used for the translation.
•Window mask register 0–3 (Wn_MASK)—defines the size of the window.
•Translated base register (Tn_BASE)—holds the base address used to
relocate the PCI address in the 21164’s memory space (for direct mapping)
and is also used to hold the scatter-gather map base address for scattergather mapping.
3–12 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
Wn_BASE[Wn_BASE_SG] determines how the PCI address is translated as
follows:
•IfWn_BASE[Wn_BASE_SG] is clear, the address is directly mapped by
concatenating Tn_BASE[CSR_WR_DATA] with the address.
•IfWn_BASE[Wn_BASE_SG] is set, the address is mapped through the
scatter-gather table, allowing any 8KB PCI address to map to any 8KB
memory address.
Figure 3–3 illustrates the mapping. This scatter-gather table is located in
memory, but the CIA provides an 8-entry TLB that caches the most recent
scatter-gather table entries. Each TLB entry holds four consecutive scattergather table entries, mapping a contiguous 32KB of virtual PCI addresses to
any four 8KB memory pages.
Figure 3–3 Address Space Overview
21164 CPU
Cached Memory Space (8GB)
8KB
Page
PCI Window
Direct Map
PCI Window
Scatter-Gather
Map
PCI Memory
LJ-04220.AI
Preliminary Edition—Subject to Change—September 1995 3–13
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
3.2.6 DMA Read Transaction
The translated address stored in the PA register is sent to the 21164 through
the flush/read address register and to memory through the memory port. If
the 21164 data cache contains valid modified data, then a copy is sent to the
Bcache-buffer part of one of the two DMA buffers in the DSW and the valid
bit is set. Memory data is always fetched and put in the memory buffer of the
DMA buffer. The PCI-buffer part of the DMA buffer is not used for DMA read
transactions.
When the first valid QW is available in the DSW’s DMA buffer, it is sent to
the DMA read buffer in the CIA through the multiplexers. Any ECC correction
is done in the CIA. From the DMA buffer in the CIA, the data goes from a
register onto the PCI a cycle later. The DMA read data also takes the bypass
path around the multiplexer for as long as the PCI can accept this stream of
data. Once the PCI stalls, then the buffers are used.
3.2.7 DMA Read Transaction (Prefetch)
The PCI supports three types of memory read commands, which are used as
specified in the PCI Local Bus Specification:
•Read—used for small transfers (up to 1/2 a cache line)
•Read line—used for medium transfers (1/2 to 3 cache lines)
•Read multiple—used for large transfers (more than 3 cache lines)
In the PCI environment, most devices assume a processor cache of 32 bytes
(compared to the CIA cache line of 64 bytes). The CIA prefetch strategy is
shown in Table 3–2.
Table 3–2 CIA Prefetch Strategy
PCI Memory Read CommandPrefetch Operation
ReadNone.
Read linePrefetch one block.
Read multiplePrefetch until end of transaction.
3–14 Preliminary Edition—Subject to Change—September 1995
A counter is used to increment the memory block address for prefetching. The
output of this counter goes to the Bcache and memory like a normal DMA read
command address. The returned data is sent to the next free DMA buffer in
the DSW. So as one buffer is being copied down to the CIA, the other is free to
accept DMA prefetched data.
At the end of a transaction all prefetched data in the DMA read buffers is
ignored (that is, no prefetch caching is performed). Prefetching does not
continue over 8KB page boundaries.
The 64-byte DMA read buffer in the CIA acts like two 32-byte halves during
DMA commands—as one half is emptying to the PCI bus, the other half is
being filled.
3.2.8 DMA Write Transaction
The translated address stored in the PA register is sent to the Bcache through
the flush/read address register and also to memory via the memory port. If
the Bcache contains valid (modified) data, the data is sent to the Bcache-buffer
portion of the DMA buffer in the DSW and the Bcache data is invalidated.
Memory data is always fetched and is put in the memory-buffer part of the
DMA buffer.
The DMA write data is taken from the PCI bus and is placed in the DMA write
buffer. If the data is a complete quadword, then ECC is generated and the
data is sent to the PCI-write part of the DMA buffer in the DSW. If the data
is an incomplete quadword, then a merge operation has to be performed. The
valid Bcache or memory quadword is sent to the CIA from the DMA buffer in
the DSW, and the valid bytes of the DMA write data are merged. This merged
quadword then has ECC generated and is sent to the PCI-write buffer portion
of the DMA buffer in the DSW.
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
The DMA buffer in the DSW builds the data up to 64 bytes (the block size)
before sending the data to memory. The memory logic generates the memory
write pulses. The memory address is read from the PA register by way of a
multiplexer.
Note
Should the I/O port be busy with an I/O write/read transaction, the
memory port will be available for DMA write commands.
Preliminary Edition—Subject to Change—September 1995 3–15
DECchip 21171-CA Architecture Overview
3.2 CIA Transactions
One reason for providing a copy of the DMA write buffer in the CIA is that
the data path to the DSW may not be available at the start of a DMA write
transaction It could be busy transferring I/O write transaction data from the
DSW to the I/O write buffers if a 21164 I/O write command had been received
concurrently with a PCI DMA write command.
DMA write transaction data is always written to memory, and is never stored
(cached) in the DMA write buffers across transactions.
3.3 System Data Coherency Issues
The 21164 is the processor in systems using the DECchip 21171 Core Logic
Chipset and the data coherency requirements of the 21164 dominate CIA data
coherency issues. The CIA supports the flush-based data coherency protocol
with a Bcache but with no duplicate tag stores for either the Scache or Bcache.
Therefore the CIA uses the read and flush transactions to probe the 21164
cache.
3.3.1 Basic Properties of the System
The system has these basic properties:
1. The memory arrays do only one thing at a time.
2. A read or write transaction to memory is atomic. No operation of the
memory arrays can occur between the read and write parts of a readmodify-write sequence.
3. The memory array sequence for a 21164 cache ‘‘miss’’ with victim
transaction is atomic in the memory array. A cache miss with victim
transaction causes the 21164 to perform a read miss transaction for the
block required for the cache miss transaction and then a Bcache victim
transaction to write the victim data to memory. No other memory access
can come between the read miss transaction and the Bcache victim
transaction.
4. When ‘‘transaction A’’ gains access to memory it is said to ‘‘own’’ memory.
Once ‘‘transaction A’’ ‘‘owns’’ memory there will be no access for any other
transaction until ‘‘transaction A’’ has completed all of its memory activity.
5. When an I/O TLB ‘‘miss’’ occurs, the CIA issues a read command to the
21164 for the TLB entry and a memory access is also made for this data.
The read command is not issued to the 21164 until memory is owned for
this access.
3–16 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.3 System Data Coherency Issues
6. The CIA processes a PCI device read transaction to memory by issuing
a read command to the 21164 for the data and making a memory access
for the required data. The read command is not issued until memory is
‘‘owned’’ for the memory cycle.
7. A device write transaction to memory causes the CIA to issue a flush
transaction to the 21164 for the data and to start a memory access for the
data. The memory access serves to write data obtained from the 21164
and also serves to write the data from the PCI device to memory. The CIA
does not issue the flush command until memory is ‘‘owned’’ for the memory
write transaction.
8. We can conclude that the 2-step sequences described in 5, 6, and 7 are all
mutually atomic. In each case, the 2-step sequence is a probe of 21164 and
a memory cycle. For any pair of sequences of the types describe in 5, 6,
and 7, both steps of one of the sequences happened entirely before either of
the two steps of the other sequences.
3.3.2 Post and Run I/O Write Data Coherency
The Alpha architecture allows ‘‘posted’’ or ‘‘buffered’’ write transactions to I/O
devices, allowing the CIA to buffer write transactions from the 21164 to a PCI
device. Several read transactions from a PCI device to main memory might
be completed while write transactions to the same PCI device are buffered.
The usual way for a software programmer to ensure that the write data has
reached the device is to read a register on the same device.
The CIA performs the following operations to implement the Alpha architecture
post and run feature.
•Before a read transaction to an I/O location is processed, all preceding
21164 write transactions to devices are flushed out of the buffers.
•Before a read transaction from a PCI device to main memory is processed
all preceding write transactions by the PCI device to main memory are
flushed out of the write buffers.
Preliminary Edition—Subject to Change—September 1995 3–17
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
3.4 CIA Memory ras_h, cas_h, and Address Operation
The technique used by the CIA to eliminate ras_h cycles for victim write
transactions is described here. The method is transparent to software.
3.4.1 Traditional Mapping of 21164 Address to Memory Address
The traditional approach of mapping the 21164 address bit for bit to the
memory address is shown in Figure 3–4.
Figure 3–4 Traditional Mapping to a Memory Address
21164
Block Offset
TagIndex
Cache Address
Physical
Address
Bank SelectRowColumn
Because a cache is much smaller than available memory, multiple memory
locations will correspond to the same cache location. This correspondence is
shown in the left-hand side of Figure 3–5 where data A1, data B1, and data C1
all map to the same cache location (A1B1C1).
3–18 Preliminary Edition—Subject to Change—September 1995
Memory Address
LJ-04221.AI
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
Figure 3–5 Traditional Victim Cache to Memory Correspondence
Data A1
Data A2
Data A3
A1B1C1
Bcache
Data B1
Data B2
Data B3
Data C1
Data C2
Data C3
Typical Memory
Address Arrangement
LJ-04222.AI
Suppose that data A1 currently resides in the Bcache, but the 21164 wants to
access data C1. A cache miss will occur that will fetch data C1 and overwrite
A1 in the cache. If A1 is the only valid copy of the data (for example, a 21164
write transaction had previously updated A1), then A1 has to be written to
memory before data C1 can be written over data A1. The displaced data A1 is
referred to as a victim.
Look again at Figure 3–5 and notice that in the traditional memory addressing
scheme, data C1 and the potential victim (data A1) are well separated in
memory . They are separated by a multiple of n times the cache size. This
sparse distribution of potential victim locations means a read fill block and its
victim may not reside in the same memory page.
Preliminary Edition—Subject to Change—September 1995 3–19
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
3.4.2 CIA Mapping of 21164 Address to Memory Address
The highest tag bits are used for memory bank selection in both the traditional
and CIA address mapping schemes.
The CIA shuffles the memory address bits so that some of the high-order 21164
address bits (part of the Bcache tag) become the memory column address and
the low-order CPU address bits (part of the cache index) become the memory
row address as shown in Figure 3–6.
Figure 3–6 Modified Mapping to a Memory Address
Block Offset
TagIndex
Cache Address
21164
Physical
Address
Bank SelectRowColumn
Interchanging the memory row and column address bits, the write transaction
target and the potential victim locations reside in the same SIMMs, even
within the same row within the SIMMs, as shown in Figure 3–7. The chance
of a cache block and its victim residing in the same memory page is greatly
increased. This technique will remove all ras_h cycles from victim write
transactions. In fact, all victims will be able to share an ras_h signal with the
read miss transaction.
3–20 Preliminary Edition—Subject to Change—September 1995
Memory Address
LJ-04223.AI
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
Figure 3–7 CIA Victim Cache to Memory Correspondence
Data A1
Data B1
Data C1
A1B1C1
Bcache
Data A2
Data B2
Data C2
Data A3
Data B3
Data C3
21171 Memory
Address Arrangement
LJ-04224.AI
However, this approach hinders DMA transactions. Consecutive blocks (data
A1, data A2, and data A3), as shown in Figure 3–7, are scattered a page apart
by using this memory addressing scheme. This implies that consecutive DMA
read/write transaction blocks will require a ras_h cycle, reducing possible
bandwidth. The traditional scheme does not suffer this problem.
A compromise between the two schemes, as shown in Figure 3–8, is used by
the CIA. The four low-order bits of the index map straight to the low-order
bits of the column address (just like in a traditional scheme). The remaining
high-order index bits go to the memory row-address. The remainder of the
column address uses the tag portion of the physical address.
Preliminary Edition—Subject to Change—September 1995 3–21
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
Figure 3–8 CIA Mapping to a Memory Address
Block Offset
TagIndex
Cache Address
21164
Physical
Address
Bank SelectRowColumn
Memory Address
LJ-04225.AI
Using this compromise scheme, a DMA transaction will march through four
64-byte blocks (64 longwords), on average, before requiring a ras_h cycle. This
constitutes a large DMA transfer, minimizing the ras_h penalty.
The logic that determines if a ras_h cycle is required is located in the memory
logic block of Figure 3–2.
3–22 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
Figures 3–9 and 3–10 show the relationship between CPU/DMA address bits
<28:04> and column and row address bits to memory. The memory address
maps for the 128-bit data path and the 256-bit data path are shown in Tables
3–9 and 3–10, respectively. The user specifies row type in MBAn[ROW_TYPE]
(MBAn<2:1>).
Figure 3–9 Memory Address Maps for 128-Bit Data Path
For 128-Bit-Wide Memory:
Row
Type
Row
Type
Note:
CPU tag index requires a minimum cache size of 2MB (CPU/DMA address bits <21:7> required).
DRAM
Size
01
16MB
1064MB
DRAM
Size
01
16MB
1064MB
DRAM
Configuration
10R x 10C4MB00
11R x 11C
12R x 10C
12R x 12C
13R x 11C
DRAM
Configuration
10R x 10C4MB00
11R x 11C
12R x 10C
12R x 12C
13R x 11C
28 27 26 25 24 23 22
-
-
-
-
-
C9
C9
R1
R1
R1
-
-
C8
C7
-
-
C8
C7
-
-
C8
C7
R0
-
-
C4
R0
R0
R11
-
C10
R11
-
-
-
-
-
-
-
-
-
-
C5
R3
R3
R3
-
-
C4
-
-
C10
R2
-
-
R2
-
-
R2
-
-
-
-
-
-
-
C3
15 14 13 12 11 10 09
R5
R4
-
-
R5
R4
-
-
R5
R4
-
-
21
20
19
18
17
-
-
-
R9
R8
C6
C5
-
-
-
R10
-
R10
-
07
-
C2
-
C2
-
C2
R9
-
R9
-
06
-
C1
-
C1
-
C1
C6
-
-
C6
08
-
-
C3
-
C3
R12
-
C11
-
R8
-
R8
-
05*
-
C0
-
C0
-
C0
R7
-
R7
-
R7
-
04*
-
C9
-
C5
-
C4
LJ-04399.AI5
16
R6
-
R6
-
R6
-
Preliminary Edition—Subject to Change—September 1995 3–23
DECchip 21171-CA Architecture Overview
3.4 CIA Memory ras_h, cas_h, and Address Operation
Figure 3–10 Memory Address Maps for 256-Bit Data Path
For 256-Bit-Wide Memory:
Row
Type
Row
Type
Note:
CPU tag index requires a minimum cache size of 2MB (CPU/DMA address bits <21:7> required).
DRAM
Size
01
16MB
1064MB
DRAM
Size
01
16MB
1064MB
DRAM
Configuration
10R x 10C4MB00
11R x 11C
12R x 10C
12R x 12C
13R x 11C
DRAM
Configuration
10R x 10C4MB00
11R x 11C
12R x 10C
12R x 12C
13R x 11C
28 27 26 25 24 23 22
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
C4
C3
15 14 13 12 11 10 09
R5
R4
-
-
R5
R4
-
-
R5
R4
-
-
C5
C5
R3
R3
R3
-
C4
-
-
C10
R2
-
-
R2
-
-
R2
-
-
C9
C9
C9
R1
R1
R1
-
C8
C7
-
-
C8
C7
-
-
C8
C7
R0
-
-
C4
R0
R0
R11
-
C10
R11
-
-
-
21
20
19
18
17
-
-
-
R9
R8
C6
C5
-
-
-
R10
-
R10
-
07
-
C2
-
C2
-
C2
R9
-
R9
-
06
-
C1
-
C1
-
C1
C6
-
-
C6
08
-
-
C3
-
C3
R12
-
C11
-
R8
-
R8
-
05*
-
C0
-
C0
-
C0
R7
-
R7
-
R7
-
04*
-
-
-
-
-
-
LJ-04400.AI5
16
R6
-
R6
-
R6
-
3–24 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.5 MB and LOCK Instructions
The MB (memory barrier) and LOCK instructions are described in this section.
3.5.1 MB Instruction
The MB instruction may or may not be disabled from leaving the 21164 by
making the BC_CONTROL [EI_CMD_GRP3] bit clear. If BC_CONTROL
[EI_CMD_GRP3] is clear, the CIA should never acknowledge receipt of an MB
instruction.
CACK_EN [MB_ENABLE], in the CIA, enables or disables the CIA from
acknowledging receipt of an MB instruction from the 21164. If CACK_EN
[MB_ENABLE] is clear, the CIA will treat an MB instruction as a NOP. See
Section 3.3.2.
3.5.2 LOCK Instruction
The CIA has to contend with locks originating on either the 21164 or the
PCI. There are two cases that require that the 21164 internal lock state on
a block be cleared (from a system point of view). The first is during a DMA
write transaction to the locked block and the second is when a PCI device has
established a lock.
3.5 MB and LOCK Instructions
The things to consider for system lock behavior are:
•DMA write transaction case—The 21164 maintains its lock flag and
lock address in its bus interface unit (BIU). This lock status is correctly
maintained, even if the locked block is displaced from the Scache, as long
as the system sends all DMA write FLUSH commands to the 21164 (no
duplicate cache tag).
•PCI lock established—A PCI device can only establish a lock on a 16-byte
block of data during a DMA read transaction. Once the lock is established
by the successful DMA read transaction, the PCI device has exclusive
access to that 16-byte block (the PCI target may lock a larger block). This
PCI lock allows the device to do an atomic read-modify-write on a 16-byte
data unit.
The Alpha and PCI architectures have different mechanisms for establishing a
lock:
•For PCI architecture, it is a successful read transaction (with the PCI
LOCK signal asserted) that establishes a lock.
•For Alpha architecture, it is a successful write transaction (STn_C) that
establishes a lock.
Preliminary Edition—Subject to Change—September 1995 3–25
DECchip 21171-CA Architecture Overview
3.5 MB and LOCK Instructions
So, a PCI read transaction with the PCI LOCK signal asserted must be treated
by the 21164 as a write transaction (that is, for the Alpha architecture, it
must have the same effect as a write transaction from some other processor).
Example 3–1 shows an example of the 21164 and PCI competing for a memory
flag.
Both the 21164 and the PCI device are attempting to perform an atomic readmodify-write operation on a memory flag (a semaphore). The 21164 begins
with a LDQ_L instruction but will not know if the update is successful until
the companion STQ_C instruction completes successfully.
Assume that the PCI device performs a PCI read transaction with the PCI
LOCK signal asserted, and obtains the lock to the memory flag before the
21164 STQ_C instruction is executed. The 21164’s STQ_C instruction must
fail, otherwise, the 21164 and the PCI device will both believe they have
successfully modified the flag.
Before a PCI device establishes the PCI lock, the 21164 must clear its lock
flag. The CIA must perform a flush transaction to the locked block, causing the
21164 to clear its lock flag. Unfortunately, the system must not only perform
a DMA read transaction, but must also write the flushed block to memory.
Once the CIA successfully sends the data to the PCI device, the PCI lock is
established.
3–26 Preliminary Edition—Subject to Change—September 1995
DECchip 21171-CA Architecture Overview
3.5 MB and LOCK Instructions
While the PCI lock is in effect, the 21164 may attempt to refill the flushed
block. This will happen if a loop, as shown in Example 3–1, is repeatedly
executing an LDx_L instruction. The requested block cannot be provided until
the PCI lock has completed because once the block is in the 21164 internal
cache, the CIA loses control of the block.
Initially, the CIA could prevent the 21164 from setting its internal lock flag for
the block by using the 21164 input signal line system_lock_flag_h. The next
time the LDx_L/STn_C loop, the 21164 will have a cache hit on the block and
will be unaware of the system lock state, so the loop will succeed this second
time.
The CIA lock rules are:
•The 21164 system_lock_flag_h input signal line is not connected to the
CIA and is always tied true.
•All DMA write transactions will result in the CIA performing a flush
transaction to the 21164.
•All DMA read transactions with the PCI LOCK signal asserted will result
in the CIA performing a flush transaction to the 21164. The CIA must
write the flushed block to memory. The internal CIA lock address register
will be set with the locking address.
•If the 21164 requests a fill that hits the CIA lock address register, then the
fill is stalled until the PCI lock state is relinquished by the PCI device.
•The 21164 lock transaction will be treated as a NOP for systems with
duplicate tag stores.
•The 21164 write transaction with lock will be treated as a write
transaction.
3.5.3 Locks to Uncached Space
The CIA does not support LDx_L and STn_C to noncacheable space. They are
converted to LDx and STn.
Preliminary Edition—Subject to Change—September 1995 3–27
DECchip 21171-CA Control and Status
This chapter describes the DECchip 21171-CA (CIA) control and status
registers (CSRs). It also provides information about programming memory
timing, configuring memory, and initializing the backup cache (Bcache).
Programmers must write only zeros to register fields defined as
reserved.
4.1 DECchip 21171-CA Registers
The DECchip 21171-CA CSRs are 32 bits wide and are located on NATURALLY
ALIGNED 64-byte addresses. Write transactions to read-only registers
could result in UNPREDICTABLE behavior while read transactions are
nondestructive. Only zeros should be written to reserved bits. The register
descriptions contain the initialized states.
4
Registers
Note
The DECchip 21171-CA CSRs are presented here in seven groups:
•General registers
•Diagnostic registers
•Performance monitor registers
•Error registers
•System configuration registers
•PCI address and scatter-gather registers
•Address translation registers
Preliminary Edition—Subject to Change—September 1995 4–1
DECchip 21171-CA Control and Status Registers
4.1 DECchip 21171-CA Registers
The Alpha 21164 microprocessor CSRs, which are 64 bits wide, are described
in the Alpha 21164 Microprocessor Hardware Reference Manual. The CSRs of
most interest to uniprocessor system designers are:
•Scache control register (SC_CTL)
•Bcache control register (BC_CONTROL)
•Bcache configuration register (BC_CONFIG)
4.1.1 DECchip 21171-CA General Registers
Table 4–1 lists the DECchip 21171-CA (CIA) general registers and Section 4.2
shows and describes the individual registers.
This hardware address extension register is used to extend a PCI sparse-space
memory address up to the full 32-bit PCI address.
In sparse-addressing mode, the 21164 address provides the low-order PCI
address bits. The high-order PCI address bits <31:26> are obtained from either
the hardware address extension register or the 21164 address, depending on
sparse regions, as shown in Table 4–13.
This hardware address extension register is used to extend a PCI sparse-space
I/O address up to the full 32-bit PCI address. In sparse-addressing mode, the
21164 address provides the PCI address <24:0> and HAE_IO provides <31:25>.
At power-up this register is set to zero. In this case, sparse I/O region A and
region B both map to the lower 32MB of sparse I/O space. Setting HAE_IO to
0200 0000 will make regions A and B consecutive in the lower 64MB of PCI
I/O space.
Preliminary Edition—Subject to Change—September 1995 4–17
DECchip 21171-CA Control and Status Registers
4.2 DECchip 21171-CA General Registers
4.2.7 CIA Acknowledgment Enable Register (CACK_EN)
The CIA acknowledgment register (CACK_EN) allows software to decide if
the CIA will acknowledge receipt any of four particular commands from the
21164 or to ignore any of these commands (NOP response).
Figure 4–7 CIA Acknowledgment Enable Register (87.4000.0600)
3104 03 02 01 00
Reserved
BC_VICTIM_EN
SET_DIRTY_EN
MB_EN
LOCK_EN
Table 4–17 CIA Acknowledgment Enable Register
FieldNameTypeDescription
LJ-04232.AI
<0>LOCK_ENRW,11—Enables CIA to acknowledge receipt of a
<1>MB_ENRW,11—Enables CIA to acknowledge receipt of an MB
<2>SET_DIRTY_ENRW,11—Enables CIA to acknowledge receipt of a SET
<3>BC_VICTIM_ENRW,11—Enables CIA to acknowledge receipt of
<31:4>ReservedRO,0—
4–18 Preliminary Edition—Subject to Change—September 1995
LOCK command from the 21164.
command from 21164.
DIRTY command from the 21164.
a BC_VICTIM command from the 21164.
Loading...
+ hidden pages
You need points to download manuals.
1 point = 1 manual.
You can buy points or you can get point for every manual you upload.