On 30.01.20 12:20, David Hildenbrand wrote:
> On 30.01.20 12:14, Christian Borntraeger wrote:
>>
>>
>> On 30.01.20 12:01, Christian Borntraeger wrote:
>>>
>>>
>>> On 30.01.20 10:49, David Hildenbrand wrote:
>>>> On 30.01.20 09:55, Christian Borntraeger wrote:
>>>>> The initial CPU reset currently clobbers the userspace fpc. This was an
>>>>> oversight during a fixup for the lazy fpu reloading rework. The reset
>>>>> calls are only done from userspace ioctls. No CPU context is loaded, so
>>>>> we can (and must) act directly on the sync regs, not on the thread
>>>>> context. Otherwise the fpu restore call will restore the zeroes fpc to
>>>>> userspace.
>>>>>
>>>>> Cc: stable@xxxxxxxxxx
>>>>> Fixes: 9abc2a08a7d6 ("KVM: s390: fix memory overwrites when vx is disabled")
>>>>> Signed-off-by: Christian Borntraeger <borntraeger@xxxxxxxxxx>
>>>>> ---
>>>>> arch/s390/kvm/kvm-s390.c | 3 +--
>>>>> 1 file changed, 1 insertion(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
>>>>> index c059b86..eb789cd 100644
>>>>> --- a/arch/s390/kvm/kvm-s390.c
>>>>> +++ b/arch/s390/kvm/kvm-s390.c
>>>>> @@ -2824,8 +2824,7 @@ static void kvm_s390_vcpu_initial_reset(struct kvm_vcpu *vcpu)
>>>>> vcpu->arch.sie_block->gcr[14] = CR14_UNUSED_32 |
>>>>> CR14_UNUSED_33 |
>>>>> CR14_EXTERNAL_DAMAGE_SUBMASK;
>>>>> - /* make sure the new fpc will be lazily loaded */
>>>>> - save_fpu_regs();
>>>>> + vcpu->run->s.regs.fpc = 0;
>>>>> current->thread.fpu.fpc = 0;
>>>>> vcpu->arch.sie_block->gbea = 1;
>>>>> vcpu->arch.sie_block->pp = 0;
>>>>>
>>>>
>>>> kvm_arch_vcpu_ioctl() does a vcpu_load(vcpu), followed by the call to
>>>> kvm_arch_vcpu_ioctl_initial_reset(), followed by a vcpu_put().
>>>>
>>>> What am I missing?
>>>
>>> vcpu_load/put does no longer reload the registers lazily. We moved that out into the
>>> vcpu_run ioctl itself. (this avoids register reloading during schedule).
>>
>> see
>> e1788bb KVM: s390: handle floating point registers in the run ioctl not in vcpu_put/load
>> 31d8b8d KVM: s390: handle access registers in the run ioctl not in vcpu_put/load
>>
>> so maybe we want to change the Fixes tag to this patch.
>>
>
> Yes, because
>
> e1788bb KVM: s390: handle floating point registers in the run ioctl not
> in vcpu_put/load
>
> broke it.
>
> We should audit all users of save_fpu_regs().
I think the store status ioctl is also broken. Everything else looks sane.
