Re: [PATCH RFC] arm/kvm: report registers we failed to set

[Sebastian Ott <sebott@redhat.com>]

On Mon, 21 Jul 2025, Cornelia Huck wrote:
> If we fail migration because of a mismatch of some registers between
> source and destination, the error message is not very informative:
>
> qemu-system-aarch64: error while loading state for instance 0x0 ofdevice 'cpu'
> qemu-system-aarch64: Failed to put registers after init: Invalid argument
>
> At least try to give the user a hint which registers had a problem,
> even if they cannot really do anything about it right now.
>
> Sample output:
>
> Could not set register op0:3 op1:0 crn:0 crm:0 op2:0 to c00fac31 (is 413fd0c1)
>
> We could be even more helpful once we support writable ID registers,
> at which point the user might actually be able to configure something
> that is migratable.

That message will be even more helpful once we have the stuff in place
to actually print a reg name here (and be able to change the values for
the guest) - but I agree this would already be a great improvement over
the current message!

>
> Suggested-by: Eric Auger <eric.auger@redhat.com>
> Signed-off-by: Cornelia Huck <cohuck@redhat.com>
> ---
>
> Notes:
> - This currently prints the list of failing registers for every call to
>  write_list_to_kvmstate(), in particular for every cpu -- we might want
>  to reduce that.
> - If the macros aren't too ugly (or we manage to improve them), there
>  might be other places where they could be useful.
>
> ---
> target/arm/kvm.c | 53 ++++++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 53 insertions(+)
>
> diff --git a/target/arm/kvm.c b/target/arm/kvm.c
> index 667234485547..ac6502e0c78f 100644
> --- a/target/arm/kvm.c
> +++ b/target/arm/kvm.c
> @@ -900,6 +900,24 @@ bool write_kvmstate_to_list(ARMCPU *cpu)
>     return ok;
> }
>
> +/* pretty-print a KVM register */
> +#define CP_REG_ARM64_SYSREG_OP(_reg, _op)                       \
> +    ((uint8_t)((_reg & CP_REG_ARM64_SYSREG_ ## _op ## _MASK) >> \
> +               CP_REG_ARM64_SYSREG_ ## _op ## _SHIFT))
> +
> +#define PRI_CP_REG_ARM64_SYSREG(_reg)                    \
> +    ({                                                   \
> +        char _out[32];                                   \
> +        snprintf(_out, sizeof(_out),                     \
> +                 "op0:%d op1:%d crn:%d crm:%d op2:%d",   \
> +                 CP_REG_ARM64_SYSREG_OP(_reg, OP0),      \
> +                 CP_REG_ARM64_SYSREG_OP(_reg, OP1),      \
> +                 CP_REG_ARM64_SYSREG_OP(_reg, CRN),      \
> +                 CP_REG_ARM64_SYSREG_OP(_reg, CRM),      \
> +                 CP_REG_ARM64_SYSREG_OP(_reg, OP2));     \
> +        _out;                                            \
> +    })

Gcc correctly complains about a dangling pointer here.

Sebastian