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

[Eric Auger <eric.auger@redhat.com>]

On 8/13/25 12:01 PM, Cornelia Huck wrote:
> On Tue, Aug 12 2025, Eric Auger <eric.auger@redhat.com> wrote:
>
>> Hi Connie,
>>
>> On 7/21/25 6:19 PM, 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.
>>>
>>> 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;                                            \
>>> +    })
>>> +
>> I am afraid this is too simplistic.
>> Refering to linux/Documentation/virt/kvm/api.rst 4.68 KVM_SET_ONE_REG
>> ARM registers section
>> there are different groups of registers (upper 16b) and not all regs are
>> further identified by op0-2, crn, crm.
>> I think it would be valuable to output the group type and then the
>> formatted lower 16b, depending on the group type.
>>
>> For instance 64b ARM FW pseudo reg is formatted as
>> 0x6030 0000 0014 <regno:16>
>>
>> a diff on reg 0 results in
>> qemu-system-aarch64: Could not set register op0:0 op1:0 crn:0 crm:0
>> op2:0 to 10003 (is 10001)
>> qemu-system-aarch64: error while loading state for instance 0x0 of
>> device 'cpu'
>> qemu-system-aarch64: Could not set register op0:0 op1:0 crn:0 crm:0
>> op2:0 to 10003 (is 10001)
>> qemu-system-aarch64: Failed to put registers after init: Invalid argument
> You mean smth like
>
> Could not set FW pseudo register <regno> to <val> (is <val>)
> Could not set ID register <decoding> to <val> (is <val>)
yes

Eric
>
> ?
>