Partition Mode

ACRN is a type 1 hypervisor that supports running multiple guest operating systems (OS). Typically, the platform BIOS/bootloader boots ACRN, and ACRN loads single or multiple guest OSes. Refer to Hypervisor Startup for details on the start-up flow of the ACRN hypervisor.

ACRN supports two modes of operation: sharing mode and partition mode. This document describes ACRN’s high-level design for partition mode support.

Introduction

In partition mode, ACRN provides guests with exclusive access to cores, memory, cache, and peripheral devices. Partition mode enables developers to dedicate resources exclusively among the guests. However, there is no support today in x86 hardware or in ACRN to partition resources such as peripheral buses (e.g., PCI). On x86 platforms that support Cache Allocation Technology (CAT) and Memory Bandwidth Allocation (MBA), developers can partition Level 2 (L2) cache, Last Level Cache (LLC), and memory bandwidth among the guests. Refer to RDT Allocation Feature Supported by Hypervisor for more details on ACRN RDT high-level design and Enable Intel Resource Director Technology (RDT) Configurations for RDT configuration.

ACRN expects static partitioning of resources either by code modification for guest configuration or through compile-time config options. All the devices exposed to the guests are either physical resources or are emulated in the hypervisor. There is no need for a Device Model and Service VM. Figure 149 shows a partition mode example of two VMs with exclusive access to physical resources.

../../_images/partition-image3.png

Figure 149 Partition Mode Example with Two VMs

Guest Info

ACRN uses multi-boot info passed from the platform bootloader to know the location of each guest kernel in memory. ACRN creates a copy of each guest kernel into each of the guests’ memory. Current implementation of ACRN requires developers to specify kernel parameters for the guests as part of the guest configuration. ACRN picks up kernel parameters from the guest configuration and copies them to the corresponding guest memory.

../../_images/partition-image18.png

Figure 150 Guest Info

ACRN Setup for Guests

Cores

ACRN requires the developer to specify the number of guests and the cores dedicated for each guest. Also, the developer needs to specify the physical core used as the bootstrap processor (BSP) for each guest. As the processors are brought to life in the hypervisor, it checks if they are configured as BSP for any of the guests. If a processor is the BSP of any of the guests, ACRN proceeds to build the memory mapping for the guest, mptable, E820 entries, and zero page for the guest. As described in Guest info, ACRN creates copies of guest kernel and kernel parameters into guest memory. Figure 151 explains these events in chronological order.

../../_images/partition-image7.png

Figure 151 Event Order for Processor Setup

Memory

For each guest in partition mode, the ACRN developer specifies the size of memory for the guest and the starting address in the host physical address in the guest configuration. There is no support for HIGHMEM for partition mode guests. The developer needs to take care of two aspects for assigning host memory to the guests:

  1. Sum of guest PCI hole and guest “System RAM” is less than 4GB.

  2. Pick the starting address in the host physical address and the size so that it does not overlap with any reserved regions in host E820.

ACRN creates EPT mapping for the guest between GPA (0, memory size) and HPA (starting address in guest configuration, memory size).

E820 and Zero Page Info

A default E820 is used for all the guests in partition mode. This table shows the reference E820 layout. Zero page is created with this E820 info for all the guests.

RAM

0 - 0xEFFFFH

RESERVED (MPTABLE)

0xF0000H - 0x100000H

RAM

0x100000H - LOWMEM

RESERVED

PCI HOLE

RESERVED

Platform Info - mptable

ACRN, in partition mode, uses mptable to convey platform info to each guest. Using this platform information, number of cores used for each guest, and whether the guest needs devices with INTX, ACRN builds mptable and copies it to the guest memory. In partition mode, ACRN uses physical APIC IDs to pass to the guests.

I/O - Virtual Devices

Port I/O is supported for PCI device config space 0xcfc and 0xcf8, vUART 0x3f8, vRTC 0x70 and 0x71, and vPIC ranges 0x20/21, 0xa0/a1, and 0x4d0/4d1. MMIO is supported for vIOAPIC. ACRN exposes a virtual host-bridge at BDF (Bus Device Function) 0.0:0 to each guest. Access to 256 bytes of config space for virtual host bridge is emulated.

I/O - Passthrough Devices

ACRN, in partition mode, supports passing through PCI devices on the platform. All the passthrough devices are exposed as child devices under the virtual host bridge. ACRN does not support either passing through bridges or emulating virtual bridges. Passthrough devices should be statically allocated to each guest using the guest configuration. ACRN expects the developer to provide the virtual BDF to BDF of the physical device mapping for all the passthrough devices as part of each guest configuration.

Runtime ACRN Support for Guests

ACRN, in partition mode, supports an option to pass through LAPIC of the physical CPUs to the guest. ACRN expects developers to specify if the guest needs LAPIC passthrough using guest configuration. When the guest configures vLAPIC as x2APIC, and if the guest configuration has LAPIC passthrough enabled, ACRN passes the LAPIC to the guest. The guest can access the LAPIC hardware directly without hypervisor interception. During runtime of the guest, this option differentiates how ACRN supports inter-processor interrupt handling and device interrupt handling. This will be discussed in detail in the corresponding sections.

../../_images/partition-image16.png

Figure 152 LAPIC Passthrough

Guest SMP Boot Flow

The core APIC IDs are reported to the guest using mptable info. SMP boot flow is similar to sharing mode. Refer to VM Startup for guest SMP boot flow in ACRN. Partition mode guests startup is the same as the Service VM startup in sharing mode.

Inter-Processor Interrupt (IPI) Handling

Guests Without LAPIC Passthrough

For guests without LAPIC passthrough, IPIs between guest CPUs are handled in the same way as sharing mode in ACRN. Refer to Virtual Interrupt for more details.

Guests With LAPIC Passthrough

ACRN supports passthrough if and only if the guest is using x2APIC mode for the vLAPIC. In LAPIC passthrough mode, writes to the Interrupt Command Register (ICR) x2APIC MSR are intercepted. The guest writes the IPI info, including vector, and destination APIC IDs to the ICR. Upon an IPI request from the guest, ACRN does a sanity check on the destination processors programmed into the ICR. If the destination is a valid target for the guest, ACRN sends an IPI with the same vector from the ICR to the physical CPUs corresponding to the destination processor info in the ICR.

../../_images/partition-image14.png

Figure 153 IPI Handling for Guests With LAPIC Passthrough

Passthrough Device Support

Configuration Space Access

ACRN emulates Configuration Space Address (0xcf8) I/O port and Configuration Space Data (0xcfc) I/O port for guests to access PCI devices configuration space. Within the config space of a device, Base Address registers (BAR), offsets starting from 0x10H to 0x24H, provide the information about the resources (I/O and MMIO) used by the PCI device. ACRN virtualizes the BAR registers and for the rest of the config space, forwards reads and writes to the physical config space of passthrough devices. Refer to the I/O section below for more details.

../../_images/partition-image1.png

Figure 154 Configuration Space Access

DMA

ACRN developers need to statically define the passthrough devices for each guest using the guest configuration. For devices to DMA to/from guest memory directly, ACRN parses the list of passthrough devices for each guest and creates context entries in the VT-d remapping hardware. EPT page tables created for the guest are used for VT-d page tables.

I/O

ACRN supports I/O for passthrough devices with two restrictions.

  1. Supports only MMIO. Thus, this requires developers to expose I/O BARs as not present in the guest configuration.

  2. Supports only 32-bit MMIO BAR type.

As the guest PCI sub-system scans the PCI bus and assigns a Guest Physical Address (GPA) to the MMIO BAR, ACRN maps the GPA to the address in the physical BAR of the passthrough device using EPT. The following timeline chart explains how PCI devices are assigned to the guest and how BARs are mapped upon guest initialization.

../../_images/partition-image13.png

Figure 155 I/O for Passthrough Devices

Interrupt Configuration

ACRN supports both legacy (INTx) and MSI interrupts for passthrough devices.

INTx Support

ACRN expects developers to identify the interrupt line info (0x3CH) from the physical BAR of the passthrough device and build an interrupt entry in the mptable for the corresponding guest. As the guest configures the vIOAPIC for the interrupt RTE, ACRN writes the info from the guest RTE into the physical IOAPIC RTE. Upon the guest kernel request to mask the interrupt, ACRN writes to the physical RTE to mask the interrupt at the physical IOAPIC. When the guest masks the RTE in vIOAPIC, ACRN masks the interrupt RTE in the physical IOAPIC. Level triggered interrupts are not supported.

MSI Support

The guest reads/writes to the PCI configuration space to configure MSI interrupts using an address. Data and control registers are passed through to the physical BAR of the passthrough device. Refer to Configuration Space Access for details on how the PCI configuration space is emulated.

Virtual Device Support

ACRN provides read-only vRTC support for partition mode guests. Writes to the data port are discarded.

For port I/O to ports other than vPIC, vRTC, or vUART, reads return 0xFF and writes are discarded.

Interrupt Delivery

Guests Without LAPIC Passthrough

In ACRN partition mode, interrupts stay disabled after a vmexit. The processor does not take interrupts when it is executing in VMX root mode. ACRN configures the processor to take vmexit upon external interrupt if the processor is executing in VMX non-root mode. Upon an external interrupt, after sending EOI to the physical LAPIC, ACRN injects the vector into the vLAPIC of the vCPU running on the processor. Guests using a Linux kernel use vectors less than 0xECh for device interrupts.

../../_images/partition-image20.png

Figure 156 Interrupt Delivery for Guests Without LAPIC Passthrough

Guests With LAPIC Passthrough

For guests with LAPIC passthrough, ACRN does not configure vmexit upon external interrupts. There is no vmexit upon device interrupts and they are handled by the guest IDT.

Hypervisor IPI Service

ACRN needs IPIs for events such as flushing TLBs across CPUs, sending virtual device interrupts (e.g., vUART to vCPUs), and others.

Guests Without LAPIC Passthrough

Hypervisor IPIs work the same way as in sharing mode.

Guests With LAPIC Passthrough

Since external interrupts are passed through to the guest IDT, IPIs do not trigger vmexit. ACRN uses NMI delivery mode and the NMI exiting is chosen for vCPUs. At the time of NMI interrupt on the target processor, if the processor is in non-root mode, vmexit happens on the processor and the event mask is checked for servicing the events.

Debug Console

For details on how the hypervisor console works, refer to Hypervisor Console.

For a guest console in partition mode, ACRN provides an option to pass vmid as an argument to vm_console. vmid is the same as the one developers use in the guest configuration.

Guests Without LAPIC Passthrough

Works the same way as sharing mode.

Hypervisor Console

ACRN uses the TSC deadline timer to provide a timer service. The hypervisor console uses a timer on CPU0 to poll characters on the serial device. To support LAPIC passthrough, the TSC deadline MSR is passed through and the local timer interrupt is also delivered to the guest IDT. Instead of the TSC deadline timer, ACRN uses the VMX preemption timer to poll the serial device.

Guest Console

ACRN exposes vUART to partition mode guests. vUART uses vPIC to inject an interrupt to the guest BSP. If the guest has more than one core, during runtime, vUART might need to inject an interrupt to the guest BSP from another core (other than BSP). As mentioned in section Hypervisor IPI Service, ACRN uses NMI delivery mode for notifying the CPU running the BSP of the guest.