RA3	Cortex-A15 implementation

This course covers Cortex-A15 high-end ARM CPU

OBJECTIVES his course is split into 3 important parts: Cortex-A15 architecture Cortex-A15 software implementation and debug Cortex-A15 hardware implementation. Introduction to Hypervisor new privilege mode is done at the beginning of this course. The consequences on address translation is then explained, introducing the 2-stage translation. Decoupling guest OS from hardware using traps to Hypervisor is studied. The course also details the new features of the Generic Interrupt Controller v2, explaining how physical interrupt requests can be virtualized. The course details the new approach regarding integrated timers / counters. AXI v4 new capabilities are highlighted with regard to AXI v3. Through sequences involving a Cortex-A15 and a Cortex-A7, the hardware coherency is studied, explaining how snoop requests can be forwarded by CCI-400 interconnect. Implementation of I/O MMU-400 is also covered. Labs are run under RVDS. A more detailed course description is available on request at info@ac6-training.com

PREREQUISITES AND RELATED COURSES
	Knowledge of Cortex-A9.
	More than 12 correct answers to Cortex-A prerequisites questionnaire.
	Related courses:
		Programming with RVDS IDE,reference RV0
		VFP programming, reference RC0
		NEON programming, reference RC1 .

Outline

First day OVERVIEW OF CORTEX-A15MP Cortex-A15 architecture Organization of a SoC based on Cortex-A15MP AMBA4 coherent interconnect capabilities Inner Shareable vs Outer Shareable attribute I/O MMU 64-Byte cacheline size, integrated L2 cache VFPv4 and SIMDv2 Highlighting differences between Cortex-A9 and Cortex-A15 INSTRUCTION PIPELINE Global organization, triple issue capability Fetch / decode / rename / dispatch stages Loop mode Execution clusters Out-of-order execution, 40-entry dispatch queue Branch accelerators INTRODUCTION TO HYPERVISOR STATE Processor privilege levels state machine, user, guest OS, hypervisor Detailing the various operation modes (Bare-Metal, Hypervisor kernel and user task, Hypervisor with Guest partition) Asymmetric approach, no support for Virtualization of Secure state functionality SVC, HVC and SMC instructions Objective of the Hypervisor Hypervisor related instructions and registers List of registers that have to be saved / restored to be able to suspend / resume a guest partition Accessing banked registers or any Non-Secure mode while running in Hypervisor mode EXCEPTION MECHANISM Hypervisor vector table Utilization of Vector #5 to trap Guest partition events System Call into Hypervisor mode Asynchronous exceptions Virtual Interrupt and Abort bits control, IRQ, FIQ, external abort routing control Hypervisor exception return Taking exceptions into Hypervisor mode GENERIC INTERRUPT CONTROLLER (GICv2) Integration in a SoC based on Cortex-A15MP and Cortex-A7MP Highlighting the new features with regard to Cortex-A9MP Steering interrupts to guest OS or Hypervisor Virtual CPU interface Split EOI functionality Deactivating an interrupt source from the Virtual CPU interface Front-end interface accessed by the Guest Kernel Back-end interface accessed by the Hypervisor Second day VIRTUALIZATION EXTENSIONS New Intermediate Physical Address, 2-stage address translation Memory translation system Memory management when running in hypervisor mode Virtual Machine Identification Exposing the MMU to Other Masters, IO MMU Emulation support, trapping load and store and executing them in Hypervisor state Second-stage access permissions and attributes LARGE PHYSICAL ADDRESS EXTENSIONS SPECIFICATION (LPAE) Need to introduce support for a second stage of translation as part of the Virtualization Extensions New 3-level translation Level-1 table descriptor format Level-2 table descriptor format Attribute and Permission fields in the translation tables Improving the caching of translation entries by providing contiguous hints complete set of cache allocation hints Handling of the ASID in the LPAE New cache and TLB maintenance operations MMU IMPLEMENTATION TLB organization, L1-TLB, L2-TLB TLB match process Coherent table walk Determining the exact cause of aborts through status registers Behavior when MMU is disabled OS SUPPORT – SYNCHRONIZATION OVERVIEW Inter-Processor Interrupts Barriers Cluster ID Exclusive access monitor, implementing Boolean semaphores Global monitor Spin-lock implementation Using events Indicating the effect of Multi Core on debug interfaces Third day

LEVEL ONE SUBSYSTEM Physically Indexed Physically Tagged caches LRU replacement algorithm, implementation with a 2-way cache Speculative accesses Hit Under Miss, Miss under Miss Write streaming threshold definition Uploading the contents of L1 caches through dedicated CP15 registers MESI data cacheline states LEVEL TWO SUBSYSTEM Cache organization Strictly enforced inclusion property with L1 data caches, simplification of snooping Optional ECC / parity protection Impact of registers slices on performance L2 prefetch engine Table walk access prefetch ACE master interface ACP slave interface By means of sequences involving a multi-core Cortex-A15 and external masters, understanding how snoop requests can be used to maintain coherency of data between caches and memory GENERIC TIMER ARM generic 64-bit timers for each processor Virtual time vs Physical time Effect of virtualization on these timers Event stream purpose Kernel event stream generation Hypervisor event stream generation Gray count timer distribution scheme PERFORMANCE MONITORING VIRTUALIZATION EXTENSIONS Hypervisor performance monitoring Guest OS performance monitoring Lazy switching of PMU state by a hypervisor Reducing the number of counters available to a Guest OS Fully virtualizing the PMU identity registers Fourth day AMBA4 AXI-4 Quality of Service signaling Updated meaning of Read Allocate and Write Allocate Transaction buffering AXI-4 stream protocol Byte types, data, position, null Byte stream Sparse stream Data merging, packing, and width conversion AXI-4 lite Burst length of 1 No exclusive access support AXI Coherency Extension (ACE) Shareability domains Coherency model, cache states Additional channel signals New channels, snoop address, snoop response, snoop data Studying through sequences how a load request and a store request will be handled whenever they are marked as outer shareable requests Using ReadUnique, CleanUnique and MakeUnique requests Distributed Virtual Memory (DVM) DVM synchronization message Selecting the coherency state machine: MESI or MOESI according to the capabilities of the interconnect Snoop filtering Exported barriers DMB / DSB inner shareable, outer shareable or system HARDWARE IMPLEMENTATION Clock domains Resets, power-on reset timing diagram Valid reset combinations Power domains Power-on reset sequence, soft reset sequence Power management Maintaining coherency while CPUs are in standby state Interface to the Power Management Unit Powering down a CPU External debug over power down CCI-400 CACHE COHERENT INTERCONNECT AMBA 4 snoop request transport Snoop connectivity and control By means of sequences involving a multi-core Cortex-A15 and external masters, understanding how snoop requests can be used to maintain coherency of data between caches and memory Connecting 2 CPUs through CCI, managing coherency domains Example of Cortex-A7 dual core and Cortex-A15 dual core CORESIGHT DEBUG Program Trace Macrocell Cross Trigger Interface and Criss Trigger Matrix for multi-processor debugging Adding Virtual Machine ID in the criterion used to set a breakpoint / watchpoint Tracking VMID change in trace output

 

Duration	Cost	Calendar
4 days	0 €