Re: [kvm-unit-tests PATCH v1 2/3] lib: arm64: Add bare-metal guest execution framework

[Marc Zyngier <maz@kernel.org>]

On Mon, 16 Mar 2026 22:43:48 +0000,
Jing Zhang <jingzhangos@google.com> wrote:
> 
> To test advanced KVM features such as nested virtualization (NV) and
> GICv4 direct interrupt injection, kvm-unit-tests needs the ability to
> act as an L1 hypervisor running at EL2 and manage its own L2 guests.
> 
> Introduce a lightweight guest management library that provides the
> infrastructure to create, configure, and execute nested guests.
> 
> This framework includes:
> - Guest lifecycle management: `guest_create()` and `guest_destroy()`
>   APIs to allocate guest context and setup Stage-2 identity mappings
>   for code and stack using the s2mmu library.
> - Context switching: The `guest_run()` assembly routine handles
>   saving the host (L1) callee-saved registers and loading the guest
>   (L2) GPRs and EL1 system registers.
> - VM-Exit handling: Installs an EL2 trap handler (`guest_hyp_vectors`)
>   to intercept guest exits and route them to `guest_c_exception_handler`
>   to determine whether to return to the host test logic or resume.
> - Guest-internal exceptions: Provides `guest_el1_vectors` to catch
>   Sync, IRQ, FIQ, and SError exceptions occurring entirely within the
>   guest (EL1) without trapping to the host.
> 
> Signed-off-by: Jing Zhang <jingzhangos@google.com>
> ---
>  arm/Makefile.arm64     |   2 +
>  lib/arm64/asm/guest.h  | 156 ++++++++++++++++++++++++
>  lib/arm64/guest.c      | 197 ++++++++++++++++++++++++++++++
>  lib/arm64/guest_arch.S | 263 +++++++++++++++++++++++++++++++++++++++++
>  4 files changed, 618 insertions(+)
>  create mode 100644 lib/arm64/asm/guest.h
>  create mode 100644 lib/arm64/guest.c
>  create mode 100644 lib/arm64/guest_arch.S
> 
> diff --git a/arm/Makefile.arm64 b/arm/Makefile.arm64
> index 5e50f5ba..9026fd71 100644
> --- a/arm/Makefile.arm64
> +++ b/arm/Makefile.arm64
> @@ -41,6 +41,8 @@ cflatobjs += lib/arm64/processor.o
>  cflatobjs += lib/arm64/spinlock.o
>  cflatobjs += lib/arm64/gic-v3-its.o lib/arm64/gic-v3-its-cmd.o
>  cflatobjs += lib/arm64/stage2_mmu.o
> +cflatobjs += lib/arm64/guest.o
> +cflatobjs += lib/arm64/guest_arch.o
>  
>  ifeq ($(CONFIG_EFI),y)
>  cflatobjs += lib/acpi.o
> diff --git a/lib/arm64/asm/guest.h b/lib/arm64/asm/guest.h
> new file mode 100644
> index 00000000..1d70873d
> --- /dev/null
> +++ b/lib/arm64/asm/guest.h
> @@ -0,0 +1,156 @@
> +/*
> + * Copyright (C) 2026, Google LLC.
> + * Author: Jing Zhang <jingzhangos@google.com>
> + *
> + * SPDX-License-Identifier: LGPL-2.0-or-later
> + */
> +#ifndef _ASMARM64_GUEST_H_
> +#define _ASMARM64_GUEST_H_
> +
> +/* Offsets for assembly (Must match struct guest) */
> +#define GUEST_X_OFFSET			0
> +#define GUEST_ELR_OFFSET		248
> +#define GUEST_SPSR_OFFSET		256
> +#define GUEST_HCR_OFFSET		264
> +#define GUEST_VTTBR_OFFSET		272
> +#define GUEST_SCTLR_OFFSET		280
> +#define GUEST_VBAR_OFFSET		288
> +#define GUEST_SP_EL1_OFFSET		296
> +#define GUEST_ESR_OFFSET		304
> +#define GUEST_FAR_OFFSET		312
> +#define GUEST_HPFAR_OFFSET		320
> +#define GUEST_EXIT_CODE_OFFSET		328
> +#define GUEST_TPIDR_EL1_OFFSET		336
> +#define GUEST_ICH_VMCR_EL2_OFFSET	344

Don't hardcode offsets. Generate them.

> +
> +#ifndef __ASSEMBLY__
> +
> +#include <libcflat.h>
> +#include <asm/stage2_mmu.h>
> +
> +/* HCR_EL2 Definitions */
> +#define HCR_VM		(1UL << 0)	/* Virtualization Enable */
> +#define HCR_FMO		(1UL << 3)	/* Physical FIQ Routing */
> +#define HCR_IMO		(1UL << 4)	/* Physical IRQ Routing */
> +#define HCR_AMO		(1UL << 5)	/* Physical SError Interrupt Routing */
> +#define HCR_RW		(1UL << 31)	/* Execution State: AArch64 */
> +#define HCR_DC		(1UL << 12)	/* Default Cacheable */
> +#define HCR_E2H		(1UL << 34)	/* EL2 Host */

Please consider importing the kernel's sysreg definition, or generate
them from an official source (the architecture JSON file, for
example).

> +
> +#define HCR_GUEST_FLAGS (HCR_VM | HCR_FMO | HCR_IMO | HCR_AMO | HCR_RW | \
> +			 HCR_DC | HCR_E2H)

Just to set expectations: HCR_EL2.DC is not supported by KVM, and
likely never will. I'm hopeful that this bit (and a few others) will
eventually be deprecated because it serves no purpose. If you need a
1:1 S1 mapping, create it using (surprise!) page tables.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.