On 7/23/21 10:26 AM, Halil Pasic wrote:
On Thu, 22 Jul 2021 09:09:26 -0400
Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
On 7/21/21 10:45 AM, Halil Pasic wrote:
On Mon, 19 Jul 2021 15:35:03 -0400
Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
It was pointed out during an unrelated patch review that locks should not
[..]
-static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev)
+static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev,
+ struct kvm *kvm)
{
- /*
- * If the KVM pointer is in the process of being set, wait until the
- * process has completed.
- */
- wait_event_cmd(matrix_mdev->wait_for_kvm,
- !matrix_mdev->kvm_busy,
- mutex_unlock(&matrix_dev->lock),
- mutex_lock(&matrix_dev->lock));
-
- if (matrix_mdev->kvm) {
We used to check if matrix_mdev->kvm is null, but ...
- matrix_mdev->kvm_busy = true;
- mutex_unlock(&matrix_dev->lock);
-
- if (matrix_mdev->kvm->arch.crypto.crycbd) {
- down_write(&matrix_mdev->kvm->arch.crypto.pqap_hook_rwsem);
- matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
- up_write(&matrix_mdev->kvm->arch.crypto.pqap_hook_rwsem);
-
- kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
- }
+ if (kvm->arch.crypto.crycbd) {
... now we just try to dereference it. And ..
We used to check matrix_mdev->kvm, now the kvm pointer is passed into
the function; however, having said that, the pointer passed in should be
checked before de-referencing it.
+ down_write(&kvm->arch.crypto.pqap_hook_rwsem);
+ kvm->arch.crypto.pqap_hook = NULL;
+ up_write(&kvm->arch.crypto.pqap_hook_rwsem);
+ mutex_lock(&kvm->lock);
mutex_lock(&matrix_dev->lock);
+
+ kvm_arch_crypto_clear_masks(kvm);
vfio_ap_mdev_reset_queues(matrix_mdev->mdev);
- kvm_put_kvm(matrix_mdev->kvm);
+ kvm_put_kvm(kvm);
matrix_mdev->kvm = NULL;
- matrix_mdev->kvm_busy = false;
- wake_up_all(&matrix_mdev->wait_for_kvm);
+
+ mutex_unlock(&kvm->lock);
+ mutex_unlock(&matrix_dev->lock);
}
}
[..]
@@ -1363,14 +1323,11 @@ static void vfio_ap_mdev_release(struct mdev_device *mdev)
{
struct ap_matrix_mdev *matrix_mdev = mdev_get_drvdata(mdev);
- mutex_lock(&matrix_dev->lock);
- vfio_ap_mdev_unset_kvm(matrix_mdev);
- mutex_unlock(&matrix_dev->lock);
-
.. before access to the matrix_mdev->kvm used to be protected by
the matrix_dev->lock ...
vfio_unregister_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
&matrix_mdev->iommu_notifier);
vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
&matrix_mdev->group_notifier);
+ vfio_ap_mdev_unset_kvm(matrix_mdev, matrix_mdev->kvm);
... but it is not any more. BTW I don't think the code is guaranteed
to fetch ->kvm just once.
There are a couple of things to point out here:
1. The vfio_ap_mdev_unset_kvm function() is the only place where the
matrix_mdev->kvm pointer is cleared. That function is called here
as well as by the group notifier callback for VFIO_GROUP_NOTIFY_SET_KVM
events. If you notice in the code above, the group notifier is
unregistered
before calling the unset function, so either the notifier will have
already
been invoked and the pointer cleared (which is why you are correct
that the KVM pointer passed in needs to get checked in the unset
function),
or will get cleared here.
Hm, vfio_unregister_notifier() indeed seems to guarantee, that by the
time it returns no notifer is running. I didn't know that. But this
blocking notifier chain uses an rwsem. So mutual exclusion with
vfio_ap_mdev_open() is guaranteed, than it is indeed guaranteed. A quick
glance at the code didn't tell me if vfio/mdev guarantees that.
I mean it would make sense to me to make the init and the cleanup
mutually exclusive, but I'm reluctant to just assume it is like that.
Can you please point me into the right direction?
I'm not quite sure what you're asking for here, but I'll give it a
shot. The notifier is registered by the vfio_ap_mdev_open()
callback which is invoked in response to the opening of the mdev fd.
Since mdev fd can't be closed unless and until it's open,
there is no way for the vfio_ap_mdev_release() callback, which
is invoked when the mdev fd is closed, to execute at the same
time as the vfio_ap_mdev_open callback. Since the release
callback unregisters the notifier and both the registration and
unregistration are done under the same rwsem, I don't see how
they can be anything but mutually exclusive. Am I missing something
here?
2. The release callback is invoked when the mdev fd is closed by userspace.
The remove callback is the only place where the matrix_mdev is
freed. The
remove callback is not called until the mdev fd is released, so it
is guaranteed
the matrix_mdev will exist when the release callback is invoked.
3. The matrix_dev->lock is then taken in the vfio_ap_mdev_unset_kvm function
before doing any operations that modify the matrix_mdev.
Yeah but both the reader, and the writer needs to use the same lock to
have the protected by the lock type of situation. That is why I asked
about the place where you read matrix_mdev members outside the
matrix_dev->lock.
With this patch, the matrix_mdev is always written or read with
the matrix_dev->lock mutex locked. By moving the locking of the
kvm->lock mutex out of the functions that set/clear the masks
in the guest's CRYCB, we are now able to lock the kvm->lock
mutex before locking the matrix_dev->lock mutex in both the
vfio_ap_mdev_set_kvm() and vfio_ap_mdev_unset_kvm()
functions. This precludes the need to unlock the
matrix_dev->lock mutex while the masks are being set or
cleared and alleviates the lockdep splat for which the open coded
locks were created.
Regards,
Halil
