On 2/14/22 19:50, Tony Krowiak wrote:
Let's allow adapters, domains and control domains to be hot plugged
into and hot unplugged from a KVM guest using a matrix mdev when an
adapter, domain or control domain is assigned to or unassigned from
the matrix mdev.
Whenever an assignment or unassignment of an adapter, domain or control
domain is performed, the AP configuration assigned to the matrix
mediated device will be filtered and assigned to the AP control block
(APCB) that supplies the AP configuration to the guest so that no
adapter, domain or control domain that is not in the host's AP
configuration nor any APQN that does not reference a queue device bound
to the vfio_ap device driver is assigned.
After updating the APCB, if the mdev is in use by a KVM guest, it is
hot plugged into the guest to dynamically provide access to the adapters,
domains and control domains provided via the newly refreshed APCB.
Keep in mind that the matrix_dev->guests_lock must be taken outside of the
matrix_mdev->kvm->lock which in turn must be taken outside of the
matrix_dev->mdevs_lock in order to avoid circular lock dependencies (i.e.,
a lockdep splat).Consequently, the locking order for hot plugging the
guest's APCB must be:
matrix_dev->guests_lock => matrix_mdev->kvm->lock => matrix_dev->mdevs_lock
Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxx>
---
drivers/s390/crypto/vfio_ap_ops.c | 198 +++++++++++++++++++-----------
1 file changed, 125 insertions(+), 73 deletions(-)
diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
index 623a4b38676d..4c382cd3afc7 100644
--- a/drivers/s390/crypto/vfio_ap_ops.c
+++ b/drivers/s390/crypto/vfio_ap_ops.c
@@ -317,10 +317,25 @@ static void vfio_ap_matrix_init(struct ap_config_info *info,
matrix->adm_max = info->apxa ? info->Nd : 15;
}
-static void vfio_ap_mdev_filter_cdoms(struct ap_matrix_mdev *matrix_mdev)
+static void vfio_ap_mdev_hotplug_apcb(struct ap_matrix_mdev *matrix_mdev)
{
+ if (matrix_mdev->kvm)
+ kvm_arch_crypto_set_masks(matrix_mdev->kvm,
+ matrix_mdev->shadow_apcb.apm,
+ matrix_mdev->shadow_apcb.aqm,
+ matrix_mdev->shadow_apcb.adm);
+}
This function updates a kvm guest's apcb. So let's rename it to
vfio_ap_update_apcb(). You can also call this function in vfio_ap_mdev_set_kvm,
instead of duplicating the code to call kvm_arch_crypto_set_masks().
+static bool vfio_ap_mdev_filter_cdoms(struct ap_matrix_mdev *matrix_mdev)
+{
+ DECLARE_BITMAP(shadow_adm, AP_DOMAINS);
+
+ bitmap_copy(shadow_adm, matrix_mdev->shadow_apcb.adm, AP_DOMAINS);
bitmap_and(matrix_mdev->shadow_apcb.adm, matrix_mdev->matrix.adm,
(unsigned long *)matrix_dev->info.adm, AP_DOMAINS);
+
+ return !bitmap_equal(shadow_adm, matrix_mdev->shadow_apcb.adm,
+ AP_DOMAINS);
}
your variable, shadow_adm, should be named original_adm. Since it represents
the original value before filtering. This makes the intent much more clear.
Same goes for the vars in vfio_ap_mdev_filter_matrix().
...
+/**
+ * vfio_ap_mdev_get_locks - acquire the locks required to assign/unassign AP
+ * adapters, domains and control domains for an mdev in
+ * the proper locking order.
+ *
+ * @matrix_mdev: the matrix mediated device object
+ */
+static void vfio_ap_mdev_get_locks(struct ap_matrix_mdev *matrix_mdev)
+{
+ /* Lock the mutex required to access the KVM guest's state */
+ mutex_lock(&matrix_dev->guests_lock);
+
+ /* If a KVM guest is running, lock the mutex required to plug/unplug the
+ * AP devices passed through to the guest
+ */
+ if (matrix_mdev->kvm)
+ mutex_lock(&matrix_mdev->kvm->lock);
+
+ /* The lock required to access the mdev's state */
+ mutex_lock(&matrix_dev->mdevs_lock);
+}
Simplifying the cdoe, and removing duplication by moving the locking code to a
function is probably a good thing. But I don't feel like this belongs to this
particular patch. In general, a patch should only do one thing, and ideally that
one thing should be as small as reasonably possible. This makes the patch easier
to read and to review.
I feel like, as much as possible, you should refactor the locking in a series
of patches that are all kept together. Ideally, they would be a patch series
completely separate from dynamic ap. After all, this series is already at 18
patches. :)
...
/**
* assign_adapter_store - parses the APID from @buf and sets the
* corresponding bit in the mediated matrix device's APM
@@ -649,17 +723,9 @@ static ssize_t assign_adapter_store(struct device *dev,
int ret;
unsigned long apid;
DECLARE_BITMAP(apm, AP_DEVICES);
-
struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev);
- mutex_lock(&matrix_dev->guests_lock);
- mutex_lock(&matrix_dev->mdevs_lock);
-
- /* If the KVM guest is running, disallow assignment of adapter */
- if (matrix_mdev->kvm) {
- ret = -EBUSY;
- goto done;
- }
+ vfio_ap_mdev_get_locks(matrix_mdev);
ret = kstrtoul(buf, 0, &apid);
if (ret)
@@ -671,8 +737,6 @@ static ssize_t assign_adapter_store(struct device *dev,
}
set_bit_inv(apid, matrix_mdev->matrix.apm);
- memset(apm, 0, sizeof(apm));
- set_bit_inv(apid, apm);
ret = vfio_ap_mdev_validate_masks(matrix_mdev);
It looks like you moved the memset() and set_bit_inv() to be closer to where
"apm" is used, namely, the call to vfio_ap_mdev_filter_matrix(). Any reason you
cannot move it down under the call to vfio_ap_mdev_link_adapter()? That would
get it even closer to where it is used.
Also, I think renaming apm to apm_delta or apm_diff makes sense here. After all,
it is the difference between the original apm, and the new apm. The new apm
has an extra bit for the newly added adapter. Do I have that right? If so, I
think renaming the variable will make the code clearer.
Both of the above comments also apply to assign_domain_store().
--
-- Jason J. Herne (jjherne@xxxxxxxxxxxxx)
