Re: [PATCH v1 4/7] vfio: ap: AP Queue Interrupt Control VFIO ioctl calls

[Cornelia Huck <cohuck@redhat.com>]

On Wed, 7 Nov 2018 23:23:40 +0100
Pierre Morel <pmorel@linux.ibm.com> wrote:

> On 07/11/2018 10:46, Cornelia Huck wrote:
> > On Wed, 31 Oct 2018 19:12:54 +0100
> > Pierre Morel <pmorel@linux.ibm.com> wrote:
> >   
> >> This is the implementation of the VFIO ioctl calls to handle
> >> the AQIC interception and use GISA to handle interrupts.
> >>
> >> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
> >> ---
> >>   drivers/s390/crypto/vfio_ap_ops.c | 95 +++++++++++++++++++++++++++++++
> >>   1 file changed, 95 insertions(+)
> >>
> >> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
> >> index 272ef427dcc0..f68102163bf4 100644
> >> --- a/drivers/s390/crypto/vfio_ap_ops.c
> >> +++ b/drivers/s390/crypto/vfio_ap_ops.c
> >> @@ -895,12 +895,107 @@ static int vfio_ap_mdev_get_device_info(unsigned long arg)
> >>   	return copy_to_user((void __user *)arg, &info, minsz);
> >>   }
> >>   
> >> +static int ap_ioctl_setirq(struct ap_matrix_mdev *matrix_mdev,
> >> +			   struct vfio_ap_aqic *parm)
> >> +{
> >> +	struct aqic_gisa aqic_gisa = reg2aqic(0);
> >> +	struct kvm_s390_gisa *gisa = matrix_mdev->kvm->arch.gisa;
> >> +	struct ap_status ap_status = reg2status(0);
> >> +	unsigned long p;
> >> +	int ret = -1;
> >> +	int apqn;
> >> +	uint32_t gd;
> >> +
> >> +	apqn = (int)(parm->cmd & 0xffff);  
> > 
> > It seems you always use cmd & 0xffff only. What if there is other stuff
> > in the remaining bits of cmd? Do you plan to ignore it in any case, or
> > should you actively check that there is nothing in it?
> >   
> 
> I do not think that the ioctl interface should reflect the hardware 
> interface.
> The ioctl interface ignores the remaining bits.
> We ignore the FC because we obviously want to make a AQIC FC=3
> We ignore the T bit.
> 
> But we receive the information from the intercepting software, i.e. QEMU 
> which should I think do the checks before using the ioctl interface.

Yes, it should; but you still can't know whether it actually did...

> 
> It seemed easier to me to pass the complete registers and to ignore some 
> bits in them. In case we get any change in the future
> But we could also only pass the APQN

I'd prefer to use a well-defined structure that explicitly handles the
userspace<->kernel communication. Not that we start relying on implicit
assumptions and then things break when userspace does something
different...