@@ -103,6 +103,7 @@ static const struct attribute_group vfio_queue_attr_group = {
static struct ap_driver vfio_ap_drv = {
.probe = vfio_ap_mdev_probe_queue,
.remove = vfio_ap_mdev_remove_queue,
+ .in_use = vfio_ap_mdev_resource_in_use,
.ids = ap_queue_ids,
};
@@ -617,10 +617,32 @@ static int vfio_ap_mdev_verify_no_sharing(unsigned long *mdev_apm,
return 0;
}
+/**
+ * vfio_ap_mdev_validate_masks - verify that the APQNs assigned to the mdev are
+ * not reserved for the default zcrypt driver and
+ * are not assigned to another mdev.
+ *
+ * @matrix_mdev: the mdev to which the APQNs being validated are assigned.
+ *
+ * Return: One of the following values:
+ * o the error returned from the ap_apqn_in_matrix_owned_by_def_drv() function,
+ * most likely -EBUSY indicating the ap_perms_mutex lock is already held.
+ * o EADDRNOTAVAIL if an APQN assigned to @matrix_mdev is reserved for the
+ * zcrypt default driver.
+ * o EADDRINUSE if an APQN assigned to @matrix_mdev is assigned to another mdev
+ * o A zero indicating validation succeeded.
+ */
static int vfio_ap_mdev_validate_masks(struct ap_matrix_mdev *matrix_mdev)
{
- if (ap_apqn_in_matrix_owned_by_def_drv(matrix_mdev->matrix.apm,
- matrix_mdev->matrix.aqm))
+ int ret;
+
+ ret = ap_apqn_in_matrix_owned_by_def_drv(matrix_mdev->matrix.apm,
+ matrix_mdev->matrix.aqm);
+
+ if (ret < 0)
+ return ret;
+
+ if (ret == 1)
return -EADDRNOTAVAIL;
return vfio_ap_mdev_verify_no_sharing(matrix_mdev->matrix.apm,
@@ -711,6 +733,10 @@ static void vfio_ap_mdev_put_locks(struct ap_matrix_mdev *matrix_mdev)
* An APQN derived from the cross product of the APID being assigned
* and the APQIs previously assigned is being used by another mediated
* matrix device
+ *
+ * 5. -EAGAIN
+ * A lock required to validate the mdev's AP configuration could not
+ * be obtained.
*/
static ssize_t assign_adapter_store(struct device *dev,
struct device_attribute *attr,
@@ -897,6 +923,10 @@ static void vfio_ap_mdev_link_domain(struct ap_matrix_mdev *matrix_mdev,
* An APQN derived from the cross product of the APQI being assigned
* and the APIDs previously assigned is being used by another mediated
* matrix device
+ *
+ * 5. -EAGAIN
+ * The lock required to validate the mdev's AP configuration could not
+ * be obtained.
*/
static ssize_t assign_domain_store(struct device *dev,
struct device_attribute *attr,
@@ -1741,3 +1771,50 @@ void vfio_ap_mdev_remove_queue(struct ap_device *apdev)
kfree(q);
vfio_ap_mdev_put_qlocks(matrix_mdev);
}
+
+/**
+ * vfio_ap_mdev_resource_in_use: check whether any of a set of APQNs is
+ * assigned to a mediated device under the control
+ * of the vfio_ap device driver.
+ *
+ * @apm: a bitmap specifying a set of APIDs comprising the APQNs to check.
+ * @aqm: a bitmap specifying a set of APQIs comprising the APQNs to check.
+ *
+ * This function is invoked by the AP bus when changes to the apmask/aqmask
+ * attributes will result in giving control of the queue devices specified via
+ * @apm and @aqm to the default zcrypt device driver. Prior to calling this
+ * function, the AP bus locks the ap_perms_mutex. If this function is called
+ * while an adapter or domain is being assigned to a mediated device, the
+ * assignment operations will take the matrix_dev->guests_lock and
+ * matrix_dev->mdevs_lock then call the ap_apqn_in_matrix_owned_by_def_drv
+ * function, which also locks the ap_perms_mutex. This could result in a
+ * deadlock.
+ *
+ * To avoid a deadlock, this function will verify that the
+ * matrix_dev->guests_lock and matrix_dev->mdevs_lock are not currently held and
+ * will return -EBUSY if the locks can not be obtained.
+ *
+ * Return:
+ * * -EBUSY if the locks required by this function are already locked.
+ * * -EADDRINUSE if one or more of the APQNs specified via @apm/@aqm are
+ * assigned to a mediated device under the control of the vfio_ap
+ * device driver.
+ */
+int vfio_ap_mdev_resource_in_use(unsigned long *apm, unsigned long *aqm)
+{
+ int ret;
+
+ if (!mutex_trylock(&matrix_dev->guests_lock))
+ return -EBUSY;
+
+ if (!mutex_trylock(&matrix_dev->mdevs_lock)) {
+ mutex_unlock(&matrix_dev->guests_lock);
+ return -EBUSY;
+ }
+
+ ret = vfio_ap_mdev_verify_no_sharing(apm, aqm);
+ mutex_unlock(&matrix_dev->mdevs_lock);
+ mutex_unlock(&matrix_dev->guests_lock);
+
+ return ret;
+}
@@ -146,4 +146,6 @@ void vfio_ap_mdev_unregister(void);
int vfio_ap_mdev_probe_queue(struct ap_device *queue);
void vfio_ap_mdev_remove_queue(struct ap_device *queue);
+int vfio_ap_mdev_resource_in_use(unsigned long *apm, unsigned long *aqm);
+
#endif /* _VFIO_AP_PRIVATE_H_ */
Let's implement the callback to indicate when an APQN is in use by the vfio_ap device driver. The callback is invoked whenever a change to the apmask or aqmask would result in one or more queue devices being removed from the driver. The vfio_ap device driver will indicate a resource is in use if the APQN of any of the queue devices to be removed are assigned to any of the matrix mdevs under the driver's control. There is potential for a deadlock condition between the matrix_dev->lock used to lock the matrix device during assignment of adapters and domains and the ap_perms_mutex locked by the AP bus when changes are made to the sysfs apmask/aqmask attributes. The AP Perms lock controls access to the objects that store the adapter numbers (ap_perms) and domain numbers (aq_perms) for the sysfs /sys/bus/ap/apmask and /sys/bus/ap/aqmask attributes. These attributes identify which queues are reserved for the zcrypt default device drivers. Before allowing a bit to be removed from either mask, the AP bus must check with the vfio_ap device driver to verify that none of the queues are assigned to any of its mediated devices. The apmask/aqmask attributes can be written or read at any time from userspace, so care must be taken to prevent a deadlock with asynchronous operations that might be taking place in the vfio_ap device driver. For example, consider the following: 1. A system administrator assigns an adapter to a mediated device under the control of the vfio_ap device driver. The driver will need to first take the matrix_dev->guests_lock to potentially hot plug the adapter into the KVM guest. 2. At the same time, a system administrator sets a bit in the sysfs /sys/bus/ap/ap_mask attribute. To complete the operation, the AP bus must: a. Take the ap_perms_mutex lock to update the object storing the values for the /sys/bus/ap/ap_mask attribute. b. Call the vfio_ap device driver's in-use callback to verify that the queues now being reserved for the default zcrypt drivers are not assigned to a mediated device owned by the vfio_ap device driver. To do the verification, the in-use callback function takes the matrix_dev->guests_lock, but has to wait because it is already held by the operation in 1 above. 3. The vfio_ap device driver calls an AP bus function to verify that the new queues resulting from the assignment of the adapter in step 1 are not reserved for the default zcrypt device driver. This AP bus function tries to take the ap_perms_mutex lock but gets stuck waiting for the waiting for the lock due to step 2a above. Consequently, we have the following deadlock situation: matrix_dev->lock locked (1) ap_perms_mutex lock locked (2a) Waiting for matrix_dev->lock (2b) which is currently held (1) Waiting for ap_perms_mutex lock (3) which is currently held (2a) To prevent this deadlock scenario, the in-use callback will use mutex_trylock to take the matrix_dev->guests_lock and if the lock can not be obtained, will return -EBUSY from the function. Signed-off-by: Tony Krowiak <akrowiak@linux.ibm.com> --- drivers/s390/crypto/vfio_ap_drv.c | 1 + drivers/s390/crypto/vfio_ap_ops.c | 81 ++++++++++++++++++++++++++- drivers/s390/crypto/vfio_ap_private.h | 2 + 3 files changed, 82 insertions(+), 2 deletions(-)