Re: [RFC] Support for Arm CCA VMs on Linux

[Suzuki K Poulose]

Hi,


> We are happy to announce the early RFC version of the Arm
> Confidential Compute Architecture (CCA) support for the Linux
> stack. The intention is to seek early feedback in the following areas:
>  * KVM integration of the Arm CCA
>  * KVM UABI for managing the Realms, seeking to generalise the operations
>    wherever possible with other Confidential Compute solutions.
>    Note: This version doesn't support Guest Private memory, which will be added
>    later (see below).
>  * Linux Guest support for Realms
>

We have updated the stack for Arm CCA Linux support to RMM-v1.0-EAC2 (See links)
We are not posting the patches for review yet, as we plan to update our
stack to support the latest RMM-v1.0 specification, which includes some
functional changes to support PSCI monitoring by the VMM along with other
minor changes. All relevant components are updated on a new branch "rmm-v1.0-eac2"
Guest-mem support is not included, but is in progress.

Change log :
 - KVM RMI support updated to v1.0-eac2, with optimisations to stage2 tear down
 - Guest (Linux and kvm-unit-test) support for RSI compliant to v1.0-eac2
 - SVE, PMU support for Realms

kvmtool :
  - Dropped no-compat and switched to --loglevel (merged upstream)
  - Support for SVE, --sve-vl for vector length

> Arm CCA Introduction
> =====================
> 
> The Arm CCA is a reference software architecture and implementation that builds
> on the Realm Management Extension (RME), enabling the execution of Virtual
> machines, while preventing access by more privileged software, such as hypervisor.
> The Arm CCA allows the hypervisor to control the VM, but removes the right for
> access to the code, register state or data that is used by VM.
> More information on the architecture is available here[0].
> 
>     Arm CCA Reference Software Architecture
> 
>         Realm World    ||    Normal World   ||  Secure World  ||
>                        ||        |          ||                ||
>  EL0 x-------x         || x----x | x------x ||                ||
>      | Realm |         || |    | | |      | ||                ||
>      |       |         || | VM | | |      | ||                ||
>  ----|  VM*  |---------||-|    |---|      |-||----------------||
>      |       |         || |    | | |  H   | ||                ||
>  EL1 x-------x         || x----x | |      | ||                ||
>          ^             ||        | |  o   | ||                ||
>          |             ||        | |      | ||                ||
>  ------- R*------------------------|  s  -|---------------------
>          S             ||          |      | ||                ||
>          I             ||          |  t   | ||                ||
>          |             ||          |      | ||                || 
>          v             ||          x------x ||                ||
>  EL2    RMM*           ||              ^    ||                ||
>          ^             ||              |    ||                ||
>  ========|=============================|========================
>          |                             | SMC
>          x--------- *RMI* -------------x
> 
>  EL3                   Root World
>                        EL3 Firmware
>  ===============================================================
> Where :
>  RMM - Realm Management Monitor
>  RMI - Realm Management Interface
>  RSI - Realm Service Interface
>  SMC - Secure Monitor Call
> 
> RME introduces a new security state "Realm world", in addition to the
> traditional Secure and Non-Secure states. The Arm CCA defines a new component,
> Realm Management Monitor (RMM) that runs at R-EL2. This is a standard piece of
> firmware, verified, installed and loaded by the EL3 firmware (e.g, TF-A), at
> system boot.
> 
> The RMM provides standard interfaces - Realm Management Interface (RMI) - to the
> Normal world hypervisor to manage the VMs running in the Realm world (also called
> Realms in short). These are exposed via SMC and are routed through the EL3
> firmwre.
> The RMI interface includes:
>   - Move a physical page from the Normal world to the Realm world
>   - Creating a Realm with requested parameters, tracked via Realm Descriptor (RD)
>   - Creating VCPUs aka Realm Execution Context (REC), with initial register state.
>   - Create stage2 translation table at any level.
>   - Load initial images into Realm Memory from normal world memory
>   - Schedule RECs (vCPUs) and handle exits
>   - Inject virtual interrupts into the Realm
>   - Service stage2 runtime faults with pages (provided by host, scrubbed by RMM).
>   - Create "shared" mappings that can be accessed by VMM/Hyp.
>   - Reclaim the memory allocated for the RAM and RTTs (Realm Translation Tables)
> 
> However v1.0 of RMM specifications doesn't support:
>  - Paging protected memory of a Realm VM. Thus the pages backing the protected
>    memory region must be pinned.
>  - Live migration of Realms.
>  - Trusted Device assignment.
>  - Physical interrupt backed Virtual interrupts for Realms
> 
> RMM also provides certain services to the Realms via SMC, called Realm Service
> Interface (RSI). These include:
>  - Realm Guest Configuration.
>  - Attestation & Measurement services
>  - Managing the state of an Intermediate Physical Address (IPA aka GPA) page.
>  - Host Call service (Communication with the Normal world Hypervisor)
> 
> The specifications for the RMM software is currently at *v1.0-Beta2* and the
> latest version is available here [1].
> 
> The Trusted Firmware foundation has an implementation of the RMM - TF-RMM -
> available here [3].
> 
> Implementation
> =================
> 
> This version of the stack is based on the RMM specification v1.0-Beta0[2], with
> following exceptions :
>   - TF-RMM/KVM currently doesn't support the optional features of PMU,
>      SVE and Self-hosted debug (coming soon).
>   - The RSI_HOST_CALL structure alignment requirement is reduced to match
>      RMM v1.0 Beta1
>   - RMI/RSI version numbers do not match the RMM spec. This will be
>     resolved once the spec/implementation is complete, across TF-RMM+Linux stack.
> 
> We plan to update the stack to support the latest version of the RMMv1.0 spec
> in the coming revisions.
> 
> This release includes the following components :
> 
>  a) Linux Kernel
>      i) Host / KVM support - Support for driving the Realms via RMI. This is
>      dependent on running in the Kernel at EL2 (aka VHE mode). Also provides
>      UABI for VMMs to manage the Realm VMs. The support is restricted to 4K page
>      size, matching the Stage2 granule supported by RMM. The VMM is responsible
>      for making sure the guest memory is locked.
> 
>        TODO: Guest Private memory[10] integration - We have been following the
>        series and support will be added once it is merged upstream.
>      
>      ii) Guest support - Support for a Linux Kernel to run in the Realm VM at
>      Realm-EL1, using RSI services. This includes virtio support (virtio-v1.0
>      only). All I/O are treated as non-secure/shared.
>  
>  c) kvmtool - VMM changes required to manage Realm VMs. No guest private memory
>     as mentioned above.
>  d) kvm-unit-tests - Support for running in Realms along with additional tests
>     for RSI ABI.
> 
> Running the stack
> ====================
> 
> To run/test the stack, you would need the following components :
> 
> 1) FVP Base AEM RevC model with FEAT_RME support [4]
> 2) TF-A firmware for EL3 [5]
> 3) TF-A RMM for R-EL2 [3]
> 4) Linux Kernel [6]
> 5) kvmtool [7]
> 6) kvm-unit-tests [8]
> 
> Instructions for building the firmware components and running the model are
> available here [9]. Once, the host kernel is booted, a Realm can be launched by
> invoking the `lkvm` commad as follows:
> 
>  $ lkvm run --realm 				 \
> 	 --measurement-algo=["sha256", "sha512"] \
> 	 --disable-sve				 \

As noted above, this is no longer required.

> 	 <normal-vm-options>
> 
> Where:
>  * --measurement-algo (Optional) specifies the algorithm selected for creating the
>    initial measurements by the RMM for this Realm (defaults to sha256).
>  * GICv3 is mandatory for the Realms.
>  * SVE is not yet supported in the TF-RMM, and thus must be disabled using
>    --disable-sve
> 
> You may also run the kvm-unit-tests inside the Realm world, using the similar
> options as above.
> 
> 
> Links
> ============
> 
> [0] Arm CCA Landing page (See Key Resources section for various documentations)
>     https://www.arm.com/architecture/security-features/arm-confidential-compute-architecture
> 
> [1] RMM Specification Latest
>     https://developer.arm.com/documentation/den0137/latest
> 
> [2] RMM v1.0-Beta0 specification
>     https://developer.arm.com/documentation/den0137/1-0bet0/

 EAC2 spec: https://developer.arm.com/documentation/den0137/1-0eac2/
> 
> [3] Trusted Firmware RMM - TF-RMM
>     https://www.trustedfirmware.org/projects/tf-rmm/
>     GIT: https://git.trustedfirmware.org/TF-RMM/tf-rmm.git
> 
> [4] FVP Base RevC AEM Model (available on x86_64 / Arm64 Linux)
>     https://developer.arm.com/Tools%20and%20Software/Fixed%20Virtual%20Platforms
> 
> [5] Trusted Firmware for A class
>     https://www.trustedfirmware.org/projects/tf-a/
> 
> [6] Linux kernel support for Arm-CCA
>     https://gitlab.arm.com/linux-arm/linux-cca
>     Host Support branch:	cca-host/rfc-v1

Update branch : cca-host/rmm-v1.0-eac2

>     Guest Support branch:	cca-guest/rfc-v1

Update branch : cca-guest/rmm-v1.0-eac2

Combined tree for host and guest is also available at: "cca-full/rmm-v1.0-eac2"

> 
> [7] kvmtool support for Arm CCA
>     https://gitlab.arm.com/linux-arm/kvmtool-cca cca/rfc-v1

Update branch : cca/rmm-v1.0-eac2

> 
> [8] kvm-unit-tests support for Arm CCA
>     https://gitlab.arm.com/linux-arm/kvm-unit-tests-cca  cca/rfc-v1
> 

Update branch : cca/rmm-v1.0-eac2


Suzuki

> [9] Instructions for Building Firmware components and running the model, see
>     section 4.19.2 "Building and running TF-A with RME"
>     https://trustedfirmware-a.readthedocs.io/en/latest/components/realm-management-extension.html#building-and-running-tf-a-with-rme
> 
> [10] fd based Guest Private memory for KVM
>    https://lkml.kernel.org/r/20221202061347.1070246-1-chao.p.peng@xxxxxxxxxxxxxxx





Cc: Alexandru Elisei <alexandru.elisei@xxxxxxx>
Cc: Andrew Jones <andrew.jones@xxxxxxxxx>
Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
Cc: Chao Peng <chao.p.peng@xxxxxxxxxxxxxxx>
Cc: Christoffer Dall <christoffer.dall@xxxxxxx>
Cc: Fuad Tabba <tabba@xxxxxxxxxx>
Cc: Jonathan Cameron <jonathan.cameron@xxxxxxxxxx>
Cc: James Morse <james.morse@xxxxxxx>
Cc: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
Cc: Joey Gouly <Joey.Gouly@xxxxxxx>
Cc: Marc Zyngier <maz@xxxxxxxxxx>
Cc: Mark Rutland <mark.rutland@xxxxxxx>
Cc: Oliver Upton <oliver.upton@xxxxxxxxx>
Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
Cc: Quentin Perret <qperret@xxxxxxxxxx>
Cc: Sean Christopherson <seanjc@xxxxxxxxxx>
Cc: Steven Price <steven.price@xxxxxxx>
Cc: Thomas Huth <thuth@xxxxxxxxxx>
Cc: Will Deacon <will@xxxxxxxxxx>
Cc: Zenghui Yu <yuzenghui@xxxxxxxxxx>
To: linux-coco@xxxxxxxxxxxxxxx
To: kvmarm@xxxxxxxxxxxxxxx
Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
To: linux-kernel@xxxxxxxxxxxxxxx
To: kvm@xxxxxxxxxxxxxxx