User manual
Table Of Contents
- Intel® IXP2800 Network Processor
- Copyright
- Contents
- Introduction 1
- Technical Description 2
- 2.1 Overview
- 2.2 Intel XScale® Core Microarchitecture
- 2.3 Microengines
- 2.4 DRAM
- 2.5 SRAM
- 2.6 Scratchpad Memory
- 2.7 Media and Switch Fabric Interface
- 2.8 Hash Unit
- 2.9 PCI Controller
- 2.10 Control and Status Register Access Proxy
- 2.11 Intel XScale® Core Peripherals
- 2.12 I/O Latency
- 2.13 Performance Monitor
- Intel XScale® Core 3
- 3.1 Introduction
- 3.2 Features
- 3.3 Memory Management
- 3.4 Instruction Cache
- 3.5 Branch Target Buffer (BTB)
- 3.6 Data Cache
- 3.6.1 Overviews
- 3.6.2 Data Cache and Mini-Data Cache Operation
- 3.6.3 Data Cache and Mini-Data Cache Control
- 3.6.4 Reconfiguring the Data Cache as Data RAM
- 3.6.5 Write Buffer/Fill Buffer Operation and Control
- 3.7 Configuration
- 3.8 Performance Monitoring
- 3.9 Performance Considerations
- 3.9.1 Interrupt Latency
- 3.9.2 Branch Prediction
- 3.9.3 Addressing Modes
- 3.9.4 Instruction Latencies
- 3.9.4.1 Performance Terms
- 3.9.4.2 Branch Instruction Timings
- 3.9.4.3 Data Processing Instruction Timings
- 3.9.4.4 Multiply Instruction Timings
- 3.9.4.5 Saturated Arithmetic Instructions
- 3.9.4.6 Status Register Access Instructions
- 3.9.4.7 Load/Store Instructions
- 3.9.4.8 Semaphore Instructions
- 3.9.4.9 Coprocessor Instructions
- 3.9.4.10 Miscellaneous Instruction Timing
- 3.9.4.11 Thumb Instructions
- 3.10 Test Features
- 3.11 Intel XScale® Core Gasket Unit
- 3.12 Intel XScale® Core Peripheral Interface
- 3.12.1 XPI Overview
- 3.12.2 UART Overview
- 3.12.3 UART Operation
- 3.12.4 Baud Rate Generator
- 3.12.5 General Purpose I/O (GPIO)
- 3.12.6 Timers
- 3.12.7 Slowport Unit
- Microengines 4
- DRAM 5
- SRAM Interface 6
- SHaC - Unit Expansion 7
- Media and Switch Fabric Interface 8
- 8.1 Overview
- 8.2 Receive
- 8.3 Transmit
- 8.4 RBUF and TBUF Summary
- 8.5 CSIX Flow Control Interface
- 8.6 Deskew and Training
- 8.7 CSIX Startup Sequence
- 8.8 Interface to Command and Push and Pull Buses
- 8.9 Receiver and Transmitter Interoperation with Framers and Switch Fabrics
- 8.9.1 Receiver and Transmitter Configurations
- 8.9.2 System Configurations
- 8.9.2.1 Framer, Single Network Processor Ingress and Egress, and Fabric Interface Chip
- 8.9.2.2 Framer, Dual Network Processor Ingress, Single Network Processor Egress, and Fabric Interface Chip
- 8.9.2.3 Framer, Single Network Processor Ingress and Egress, and CSIX-L1 Chips for Translation and Fabric Interface
- 8.9.2.4 CPU Complex, Network Processor, and Fabric Interface Chip
- 8.9.2.5 Framer, Single Network Processor, Co-Processor, and Fabric Interface Chip
- 8.9.3 SPI-4.2 Support
- 8.9.4 CSIX-L1 Protocol Support
- 8.9.5 Dual Protocol (SPI and CSIX-L1) Support
- 8.9.6 Transmit State Machine
- 8.9.7 Dynamic De-Skew
- 8.9.8 Summary of Receiver and Transmitter Signals
- PCI Unit 9
- 9.1 Overview
- 9.2 PCI Pin Protocol Interface Block
- 9.2.1 PCI Commands
- 9.2.2 IXP2800 Network Processor Initialization
- 9.2.3 PCI Type 0 Configuration Cycles
- 9.2.4 PCI 64-Bit Bus Extension
- 9.2.5 PCI Target Cycles
- 9.2.6 PCI Initiator Transactions
- 9.2.7 PCI Fast Back-to-Back Cycles
- 9.2.8 PCI Retry
- 9.2.9 PCI Disconnect
- 9.2.10 PCI Built-In System Test
- 9.2.11 PCI Central Functions
- 9.3 Slave Interface Block
- 9.4 Master Interface Block
- 9.5 PCI Unit Error Behavior
- 9.5.1 PCI Target Error Behavior
- 9.5.1.1 Target Access Has an Address Parity Error
- 9.5.1.2 Initiator Asserts PCI_PERR_L in Response to One of Our Data Phases
- 9.5.1.3 Discard Timer Expires on a Target Read
- 9.5.1.4 Target Access to the PCI_CSR_BAR Space Has Illegal Byte Enables
- 9.5.1.5 Target Write Access Receives Bad Parity PCI_PAR with the Data
- 9.5.1.6 SRAM Responds with a Memory Error on One or More Data Phases on a Target Read
- 9.5.1.7 DRAM Responds with a Memory Error on One or More Data Phases on a Target Read
- 9.5.2 As a PCI Initiator During a DMA Transfer
- 9.5.2.1 DMA Read from DRAM (Memory-to-PCI Transaction) Gets a Memory Error
- 9.5.2.2 DMA Read from SRAM (Descriptor Read) Gets a Memory Error
- 9.5.2.3 DMA from DRAM Transfer (Write to PCI) Receives PCI_PERR_L on PCI Bus
- 9.5.2.4 DMA To DRAM (Read from PCI) Has Bad Data Parity
- 9.5.2.5 DMA Transfer Experiences a Master Abort (Time-Out) on PCI
- 9.5.2.6 DMA Transfer Receives a Target Abort Response During a Data Phase
- 9.5.2.7 DMA Descriptor Has a 0x0 Word Count (Not an Error)
- 9.5.3 As a PCI Initiator During a Direct Access from the Intel XScale® Core or Microengine
- 9.5.3.1 Master Transfer Experiences a Master Abort (Time-Out) on PCI
- 9.5.3.2 Master Transfer Receives a Target Abort Response During a Data Phase
- 9.5.3.3 Master from the Intel XScale® Core or Microengine Transfer (Write to PCI) Receives PCI_PERR_L on PCI Bus
- 9.5.3.4 Master Read from PCI (Read from PCI) Has Bad Data Parity
- 9.5.3.5 Master Transfer Receives PCI_SERR_L from the PCI Bus
- 9.5.3.6 Intel XScale® Core Microengine Requests Direct Transfer when the PCI Bus is in Reset
- 9.5.1 PCI Target Error Behavior
- 9.6 PCI Data Byte Lane Alignment
- Clocks and Reset 10
- 10.1 Clocks
- 10.2 Synchronization Between Frequency Domains
- 10.3 Reset
- 10.4 Boot Mode
- 10.5 Initialization
- Performance Monitor Unit 11
- 11.1 Introduction
- 11.2 Interface and CSR Description
- 11.3 Performance Measurements
- 11.4 Events Monitored in Hardware
- 11.4.1 Queue Statistics Events
- 11.4.2 Count Events
- 11.4.3 Design Block Select Definitions
- 11.4.4 Null Event
- 11.4.5 Threshold Events
- 11.4.6 External Input Events
- 11.4.6.1 XPI Events Target ID(000001) / Design Block #(0100)
- 11.4.6.2 SHaC Events Target ID(000010) / Design Block #(0101)
- 11.4.6.3 IXP2800 Network Processor MSF Events Target ID(000011) / Design Block #(0110)
- 11.4.6.4 Intel XScale® Core Events Target ID(000100) / Design Block #(0111)
- 11.4.6.5 PCI Events Target ID(000101) / Design Block #(1000)
- 11.4.6.6 ME00 Events Target ID(100000) / Design Block #(1001)
- 11.4.6.7 ME01 Events Target ID(100001) / Design Block #(1001)
- 11.4.6.8 ME02 Events Target ID(100010) / Design Block #(1001)
- 11.4.6.9 ME03 Events Target ID(100011) / Design Block #(1001)
- 11.4.6.10 ME04 Events Target ID(100100) / Design Block #(1001)
- 11.4.6.11 ME05 Events Target ID(100101) / Design Block #(1001)
- 11.4.6.12 ME06 Events Target ID(100110) / Design Block #(1001)
- 11.4.6.13 ME07 Events Target ID(100111) / Design Block #(1001)
- 11.4.6.14 ME10 Events Target ID(110000) / Design Block #(1010)
- 11.4.6.15 ME11 Events Target ID(110001) / Design Block #(1010)
- 11.4.6.16 ME12 Events Target ID(110010) / Design Block #(1010)
- 11.4.6.17 ME13 Events Target ID(110011) / Design Block #(1010)
- 11.4.6.18 ME14 Events Target ID(110100) / Design Block #(1010)
- 11.4.6.19 ME15 Events Target ID(110101) / Design Block #(1010)
- 11.4.6.20 ME16 Events Target ID(100110) / Design Block #(1010)
- 11.4.6.21 ME17 Events Target ID(110111) / Design Block #(1010)
- 11.4.6.22 SRAM DP1 Events Target ID(001001) / Design Block #(0010)
- 11.4.6.23 SRAM DP0 Events Target ID(001010) / Design Block #(0010)
- 11.4.6.24 SRAM CH3 Events Target ID(001011) / Design Block #(0010)
- 11.4.6.25 SRAM CH2 Events Target ID(001100) / Design Block #(0010)
- 11.4.6.26 SRAM CH1 Events Target ID(001101) / Design Block #(0010)
- 11.4.6.27 SRAM CH0 Events Target ID(001110) / Design Block #(0010)
- 11.4.6.28 DRAM DPLA Events Target ID(010010) / Design Block #(0011)
- 11.4.6.29 DRAM DPSA Events Target ID(010011) / Design Block #(0011)
- 11.4.6.30 IXP2800 Network Processor DRAM CH2 Events Target ID(010100) / Design Block #(0011)
- 11.4.6.31 IXP2800 Network Processor DRAM CH1 Events Target ID(010101) / Design Block #(0011)
- 11.4.6.32 IXP2800 Network Processor DRAM CH0 Events Target ID(010110) / Design Block #(0011)

Hardware Reference Manual 91
Intel
®
IXP2800 Network Processor
Intel XScale
®
Core
3.4.1.5 Parity Protection
The instruction cache is protected by parity to ensure data integrity. Each instruction cache word
has 1 parity bit. (The instruction cache tag is not parity protected.) When a parity error is detected
on an instruction cache access, a prefetch abort exception occurs if the Intel XScale
®
core attempts
to execute the instruction. Before servicing the exception, hardware place a notification of the error
in the Fault Status register (Coprocessor 15, register 5).
A software exception handler can recover from an instruction cache parity error. This can be
accomplished by invalidating the instruction cache and the branch target buffer and then returning
to the instruction that caused the prefetch abort exception. A simplified code example is shown in
Example 17. A more complex handler might choose to invalidate the specific line that caused the
exception and then invalidate the BTB.
If a parity error occurs on an instruction that is locked in the cache, the software exception handler
needs to unlock the instruction cache, invalidate the cache and then re-lock the code in before it
returns to the faulting instruction.
3.4.1.6 Instruction Cache Coherency
The instruction cache does not detect modification to program memory by loads, stores or actions
of other bus masters. Several situations may require program memory modification, such as
uploading code from disk.
The application program is responsible for synchronizing code modification and invalidating the
cache. In general, software must ensure that modified code space is not accessed until modification
and invalidating are completed.
To achieve cache coherence, instruction cache contents can be invalidated after code modification
in external memory is complete.
If the instruction cache is not enabled, or code is being written to a non-cacheable region, software
must still invalidate the instruction cache before using the newly-written code. This precaution
ensures that state associated with the new code is not buffered elsewhere in the processor, such as
the fetch buffers or the BTB.
Naturally, when writing code as data, care must be taken to force it completely out of the processor
into external memory before attempting to execute it. If writing into a non-cacheable region,
flushing the write buffers is sufficient precaution. If writing to a cacheable region, then the data
cache should be submitted to a Clean/Invalidate operation to ensure coherency.
Example 17. Recovering from an Instruction Cache Parity Error
; Prefetch abort handler
MCR P15,0,R0,C7,C5,0 ; Invalidate the instruction cache and branch target
; buffer
CPWAIT ; wait for effect
;
SUBS PC,R14,#4 ; Returns to the instruction that generated the
; parity error
; The Instruction Cache is guaranteed to be invalidated at this point