Re: [PATCH 3/3] kvm: s390: Fix lpsw/e spec exception ilc reporting

[Janosch Frank <frankja@linux.ibm.com>]

On 4/23/26 15:36, Heiko Carstens wrote:
> On Thu, Apr 23, 2026 at 12:36:04PM +0000, Janosch Frank wrote:
>> On early PSW specification exception the ILC has to be 0 according to
>> architecture.
>>
>> Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
>> Fixes: 48a3e950f4cee ("KVM: s390: Add support for machine checks.")
>> ---
>>   arch/s390/kvm/priv.c | 16 ++++++++++++----
>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/s390/kvm/priv.c b/arch/s390/kvm/priv.c
>> index 56e63679f9d1..9fd7d3f1d1e8 100644
>> --- a/arch/s390/kvm/priv.c
>> +++ b/arch/s390/kvm/priv.c
>> @@ -713,6 +713,7 @@ int is_valid_psw(psw_t *psw)
>>   int kvm_s390_handle_lpsw(struct kvm_vcpu *vcpu)
>>   {
>>   	psw_t *gpsw = &vcpu->arch.sie_block->gpsw;
>> +	struct kvm_s390_pgm_info info = {};
>>   	psw32_t new_psw;
>>   	u64 addr, iaddr;
>>   	int rc;
>> @@ -738,14 +739,18 @@ int kvm_s390_handle_lpsw(struct kvm_vcpu *vcpu)
>>   	gpsw->addr = new_psw.addr & ~PSW32_ADDR_AMODE;
>>   	vcpu->arch.sie_block->gbea = iaddr;
>>   
>> -	if (!is_valid_psw(gpsw))
>> -		return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
>> +	if (!is_valid_psw(gpsw)) {
>> +		info.code = PGM_SPECIFICATION;
>> +		info.flags = KVM_S390_PGM_FLAGS_ILC_VALID;
>> +		return kvm_s390_inject_prog_irq(vcpu, &info);
>> +	}
> 
> Hmm... looking at the architecture: an odd instruction address in the PSW
> does not result in an early specification exception, but a "normal"
> specification exception. is_valid_psw() however also checks for an odd
> instruction address. So I guess this is still not entirely correct.
> 
> I'm also wondering if the above implementation is correct if PER is enabled
> when the specification happens...

Well, we're hopefully slowly getting there one patch at a time.

> 
> Btw.: you may want to have consistent short descriptions for your two
> patches. That is consistently use upper or lower case for "LPSW/E".

Will fix
I'll also have to rename the bear test to lpsw since we're now testing 
way more things.