Re: [PATCH v1 6/7] vfio: ap: register guest ISC with GISA and GIB

[Pierre Morel]

On 06/11/2018 21:21, Tony Krowiak wrote:
> On 10/31/18 2:12 PM, Pierre Morel wrote:
> > Register to the GIB Alert list and retrieve the GAL_ISC
> > to pass to the GISA registration.
> >
> > Unregister on error and when clearing the interrupt.
> >
> > Signed-off-by: Pierre Morel <pmorel@xxxxxxxxxxxxx>
> > Signed-off-by: Michael Mueller <mimu@xxxxxxxxxxxxx>
> > ---
> > Â drivers/s390/crypto/vfio_ap_ops.c | 16 +++++++++++++---
> > Â 1 file changed, 13 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/s390/crypto/vfio_ap_ops.c 
> > b/drivers/s390/crypto/vfio_ap_ops.c
> > index f68102163bf4..232168797fb8 100644
> > --- a/drivers/s390/crypto/vfio_ap_ops.c
> > +++ b/drivers/s390/crypto/vfio_ap_ops.c
> > @@ -903,16 +903,20 @@ static int ap_ioctl_setirq(struct ap_matrix_mdev 
> > *matrix_mdev,
> > ÂÂÂÂÂ struct ap_status ap_status = reg2status(0);
> > ÂÂÂÂÂ unsigned long p;
> > ÂÂÂÂÂ int ret = -1;
> > -ÂÂÂ int apqn;
> > +ÂÂÂ int apqn, gal_isc;
> > ÂÂÂÂÂ uint32_t gd;
> > +ÂÂÂ gal_isc = kvm_s390_gisc_register(matrix_mdev->kvm, 
> > matrix_mdev->gisc);
> > +ÂÂÂ if (gal_isc < 0)
> > +ÂÂÂÂÂÂÂ return -EIO;
> > +
> > ÂÂÂÂÂ apqn = (int)(parm->cmd & 0xffff);
> > ÂÂÂÂÂ gd = matrix_mdev->kvm->vcpus[0]->arch.sie_block->gd;
> > ÂÂÂÂÂ if (gd & 0x01)
> > ÂÂÂÂÂÂÂÂÂ aqic_gisa.f = 1;
> > ÂÂÂÂÂ aqic_gisa.gisc = matrix_mdev->gisc;
> > -ÂÂÂ aqic_gisa.isc = GAL_ISC;
> > +ÂÂÂ aqic_gisa.isc = gal_isc;
> > ÂÂÂÂÂ aqic_gisa.ir = 1;
> > ÂÂÂÂÂ aqic_gisa.gisao = gisa->next_alert >> 4;
> > @@ -923,7 +927,11 @@ static int ap_ioctl_setirq(struct ap_matrix_mdev 
> > *matrix_mdev,
> > ÂÂÂÂÂ parm->status = ret;
> > ÂÂÂÂÂ ap_status = reg2status(ret);
> > -ÂÂÂ return (ap_status.rc) ? -EIO : 0;
> > +ÂÂÂ if (ap_status.rc) {
> > +ÂÂÂÂÂÂÂ kvm_s390_gisc_unregister(matrix_mdev->kvm, matrix_mdev->gisc);
> > +ÂÂÂÂÂÂÂ return -EIO;
> > +ÂÂÂ }
> > +ÂÂÂ return 0;
> > Â }
> > Â static int ap_ioctl_clrirq(struct ap_matrix_mdev *matrix_mdev,
> > @@ -946,6 +954,8 @@ static int ap_ioctl_clrirq(struct ap_matrix_mdev 
> > *matrix_mdev,
> > ÂÂÂÂÂ parm->status = retval;
> > ÂÂÂÂÂ ap_status = reg2status(retval);
> > +ÂÂÂ /* unregister the IAM from the GIB anyway! */
> > +ÂÂÂ kvm_s390_gisc_unregister(matrix_mdev->kvm, matrix_mdev->gisc);
>
> The case statement in patch 4 does not set mdev->gisc, so the
> presumption here is that VFIO_AP_SET_IRQ has been previously called and
> has set the value for matrix_mdev->gisc. Is it possible for
> VFIO_AP_CLEAR_IRQ to get invoked without a prior call to

right, I will check this.

However if the IRQ is not registered there is no problem as we will get 
an error when unregistering the IRQ
and unregistering the gisc 0 which was not registered will fail too.

But I think we better check before in case these functions change.

> VFIO_AP_SET_IRQ? In any case, shouldn't the GISC value be taken from
> bits 61-63 of 'parm'?

No this is not possible, the ISC field is not relevant when clearing 
interrupts.

> > ÂÂÂÂÂ return (ap_status.rc) ? -EIO : 0;
> > Â }


--
Pierre Morel
Linux/KVM/QEMU in BÃblingen - Germany