Re: [PATCH v3] s390/vfio-ap: clean up vfio_ap resources when KVM pointer invalidated

[Halil Pasic]

On Wed, 16 Dec 2020 10:58:48 +0100
Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:

> On 16.12.20 02:21, Halil Pasic wrote:
> > On Tue, 15 Dec 2020 19:10:20 +0100
> > Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
> > 
> >>
> >>
> >> On 15.12.20 11:57, Halil Pasic wrote:
> >>> On Mon, 14 Dec 2020 11:56:17 -0500
> >>> Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
> >>>
> >>>> The vfio_ap device driver registers a group notifier with VFIO when the
> >>>> file descriptor for a VFIO mediated device for a KVM guest is opened to
> >>>> receive notification that the KVM pointer is set (VFIO_GROUP_NOTIFY_SET_KVM
> >>>> event). When the KVM pointer is set, the vfio_ap driver takes the
> >>>> following actions:
> >>>> 1. Stashes the KVM pointer in the vfio_ap_mdev struct that holds the state
> >>>>    of the mediated device.
> >>>> 2. Calls the kvm_get_kvm() function to increment its reference counter.
> >>>> 3. Sets the function pointer to the function that handles interception of
> >>>>    the instruction that enables/disables interrupt processing.
> >>>> 4. Sets the masks in the KVM guest's CRYCB to pass AP resources through to
> >>>>    the guest.
> >>>>
> >>>> In order to avoid memory leaks, when the notifier is called to receive
> >>>> notification that the KVM pointer has been set to NULL, the vfio_ap device
> >>>> driver should reverse the actions taken when the KVM pointer was set.
> >>>>
> >>>> Fixes: 258287c994de ("s390: vfio-ap: implement mediated device open callback")
> >>>> Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxx>
> >>>> ---
> >>>>  drivers/s390/crypto/vfio_ap_ops.c | 29 ++++++++++++++++++++---------
> >>>>  1 file changed, 20 insertions(+), 9 deletions(-)
> >>>>
> >>>> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
> >>>> index e0bde8518745..cd22e85588e1 100644
> >>>> --- a/drivers/s390/crypto/vfio_ap_ops.c
> >>>> +++ b/drivers/s390/crypto/vfio_ap_ops.c
> >>>> @@ -1037,8 +1037,6 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
> >>>>  {
> >>>>  	struct ap_matrix_mdev *m;
> >>>>
> >>>> -	mutex_lock(&matrix_dev->lock);
> >>>> -
> >>>>  	list_for_each_entry(m, &matrix_dev->mdev_list, node) {
> >>>>  		if ((m != matrix_mdev) && (m->kvm == kvm)) {
> >>>>  			mutex_unlock(&matrix_dev->lock);
> >>>> @@ -1049,7 +1047,6 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
> >>>>  	matrix_mdev->kvm = kvm;
> >>>>  	kvm_get_kvm(kvm);
> >>>>  	kvm->arch.crypto.pqap_hook = &matrix_mdev->pqap_hook;
> >>>> -	mutex_unlock(&matrix_dev->lock);
> >>>>
> >>>>  	return 0;
> >>>>  }
> >>>> @@ -1083,35 +1080,49 @@ static int vfio_ap_mdev_iommu_notifier(struct notifier_block *nb,
> >>>>  	return NOTIFY_DONE;
> >>>>  }
> >>>>
> >>>> +static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev)
> >>>> +{
> >>>> +	kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
> >>>> +	matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
> >>>
> >>>
> >>> This patch LGTM. The only concern I have with it is whether a
> >>> different cpu is guaranteed to observe the above assignment as
> >>> an atomic operation. I think we didn't finish this discussion
> >>> at v1, or did we?
> >>
> >> You mean just this assigment:
> >>>> +	matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
> >> should either have the old or the new value, but not halve zero halve old?
> >>
> > 
> > Yes that is the assignment I was referring to. Old value will work as well because
> > kvm holds a reference to this module while in the pqap_hook.
> >  
> >> Normally this should be ok (and I would consider this a compiler bug if
> >> this is split into 2 32 bit zeroes) But if you really want to be sure then we
> >> can use WRITE_ONCE.
> > 
> > Just my curiosity: what would make this a bug? Is it the s390 elf ABI,
> > or some gcc feature, or even the C standard? Also how exactly would
> > WRITE_ONCE, also access via volatile help in this particular situation?
> 
> I think its a tricky things and not strictly guaranteed, but there is a lot
> of code that relies on the atomicity of word sizes. see for example the discussion
> here
> https://lore.kernel.org/lkml/CAHk-=wgC4+kV9AiLokw7cPP429rKCU+vjA8cWAfyOjC3MtqC4A@xxxxxxxxxxxxxx/
> 
> WRITE_ONCE will not change the guarantees a lot, but it is mostly a documentation
> that we assume atomic access here.

Thanks a lot! I've read it, and IMHO it seems to contradict the section
https://lwn.net/Articles/793253/#Store%20Tearing a little. From there, I also learned
that WRITE_ONCE (i.e. volatile access) can help, although I don't really
understand why. Of course, we don't need to be portable here, as this
is s390 only code. So we might be safe without anything -- I don't know.
I believe, if volatile were enough (under any circumstances), the C
standard wouldn't have introduced atomic types.

Regards,
Halil