]> git.baikalelectronics.ru Git - kernel.git/commitdiff
s390/vfio-ap: refresh guest's APCB by filtering AP resources assigned to mdev
authorTony Krowiak <akrowiak@linux.ibm.com>
Wed, 11 Aug 2021 20:23:05 +0000 (16:23 -0400)
committerAlexander Gordeev <agordeev@linux.ibm.com>
Tue, 19 Jul 2022 14:18:09 +0000 (16:18 +0200)
Refresh the guest's APCB by filtering the APQNs and control domain numbers
assigned to the matrix mdev.

Filtering of APQNs:
-----------------
APQNs that do not reference an AP queue device bound to the vfio_ap device
driver must be filtered from the APQNs assigned to the matrix mdev before
they can be assigned to the guest's APCB. Given that the APQNs are
configured in the guest's APCB as a matrix of APIDs (adapters) and APQIs
(domains), it is not possible to filter an individual APQN. For example,
suppose the matrix of APQNs is structured as follows:

                   APIDs
             3      4      5
        0  (3,0)  (4,0)  (5,0)
APQIs   1  (3,1)  (4,1)  (5,1)
        2  (3,2)  (4,2)  (5,2)

Now suppose APQN (4,1) does not reference a queue device bound to the
vfio_ap device driver. If we filter APID 4, the APQNs (4,0), (4,1) and
(4,2) will be removed. Similarly, if we filter domain 1, APQNs (3,1),
(4,1) and (5,1) will be removed.

To resolve this dilemma, the choice was made to filter the APID - in this
case 4 - from the guest's APCB. The reason for this design decision is
because the APID references an AP adapter which is a real hardware device
that can be physically installed, removed, enabled or disabled; whereas, a
domain is a partition within the adapter. It therefore better reflects
reality to remove the APID from the guest's APCB.

Filtering of control domains:
----------------------------
Any control domains that are not assigned to the host's AP configuration
will be filtered from those assigned to the matrix mdev before assigning
them to the guest's APCB.

Signed-off-by: Tony Krowiak <akrowiak@linux.ibm.com>
Reviewed-by: Jason J. Herne <jjherne@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
drivers/s390/crypto/vfio_ap_ops.c

index 5cc4e1c31aaeb2a4512f60944a51102a449ecf97..c9439ba2ff8504248deb45603c823ba746c723ed 100644 (file)
@@ -447,6 +447,68 @@ 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)
+{
+       bitmap_and(matrix_mdev->shadow_apcb.adm, matrix_mdev->matrix.adm,
+                  (unsigned long *)matrix_dev->info.adm, AP_DOMAINS);
+}
+
+/*
+ * vfio_ap_mdev_filter_matrix - filter the APQNs assigned to the matrix mdev
+ *                             to ensure no queue devices are passed through to
+ *                             the guest that are not bound to the vfio_ap
+ *                             device driver.
+ *
+ * @matrix_mdev: the matrix mdev whose matrix is to be filtered.
+ *
+ * Note: If an APQN referencing a queue device that is not bound to the vfio_ap
+ *      driver, its APID will be filtered from the guest's APCB. The matrix
+ *      structure precludes filtering an individual APQN, so its APID will be
+ *      filtered.
+ */
+static void vfio_ap_mdev_filter_matrix(unsigned long *apm, unsigned long *aqm,
+                                      struct ap_matrix_mdev *matrix_mdev)
+{
+       int ret;
+       unsigned long apid, apqi, apqn;
+
+       ret = ap_qci(&matrix_dev->info);
+       if (ret)
+               return;
+
+       vfio_ap_matrix_init(&matrix_dev->info, &matrix_mdev->shadow_apcb);
+
+       /*
+        * Copy the adapters, domains and control domains to the shadow_apcb
+        * from the matrix mdev, but only those that are assigned to the host's
+        * AP configuration.
+        */
+       bitmap_and(matrix_mdev->shadow_apcb.apm, matrix_mdev->matrix.apm,
+                  (unsigned long *)matrix_dev->info.apm, AP_DEVICES);
+       bitmap_and(matrix_mdev->shadow_apcb.aqm, matrix_mdev->matrix.aqm,
+                  (unsigned long *)matrix_dev->info.aqm, AP_DOMAINS);
+
+       for_each_set_bit_inv(apid, apm, AP_DEVICES) {
+               for_each_set_bit_inv(apqi, aqm, AP_DOMAINS) {
+                       /*
+                        * If the APQN is not bound to the vfio_ap device
+                        * driver, then we can't assign it to the guest's
+                        * AP configuration. The AP architecture won't
+                        * allow filtering of a single APQN, so let's filter
+                        * the APID since an adapter represents a physical
+                        * hardware device.
+                        */
+                       apqn = AP_MKQID(apid, apqi);
+
+                       if (!vfio_ap_mdev_get_queue(matrix_mdev, apqn)) {
+                               clear_bit_inv(apid,
+                                             matrix_mdev->shadow_apcb.apm);
+                               break;
+                       }
+               }
+       }
+}
+
 static int vfio_ap_mdev_probe(struct mdev_device *mdev)
 {
        struct ap_matrix_mdev *matrix_mdev;
@@ -802,6 +864,8 @@ static ssize_t assign_adapter_store(struct device *dev,
 {
        int ret;
        unsigned long apid;
+       DECLARE_BITMAP(apm_delta, AP_DEVICES);
+
        struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev);
 
        mutex_lock(&matrix_dev->lock);
@@ -837,6 +901,10 @@ static ssize_t assign_adapter_store(struct device *dev,
                goto share_err;
 
        vfio_ap_mdev_link_adapter(matrix_mdev, apid);
+       memset(apm_delta, 0, sizeof(apm_delta));
+       set_bit_inv(apid, apm_delta);
+       vfio_ap_mdev_filter_matrix(apm_delta,
+                                  matrix_mdev->matrix.aqm, matrix_mdev);
        ret = count;
        goto done;
 
@@ -905,6 +973,10 @@ static ssize_t unassign_adapter_store(struct device *dev,
 
        clear_bit_inv((unsigned long)apid, matrix_mdev->matrix.apm);
        vfio_ap_mdev_unlink_adapter(matrix_mdev, apid);
+
+       if (test_bit_inv(apid, matrix_mdev->shadow_apcb.apm))
+               clear_bit_inv(apid, matrix_mdev->shadow_apcb.apm);
+
        ret = count;
 done:
        mutex_unlock(&matrix_dev->lock);
@@ -978,6 +1050,7 @@ static ssize_t assign_domain_store(struct device *dev,
 {
        int ret;
        unsigned long apqi;
+       DECLARE_BITMAP(aqm_delta, AP_DOMAINS);
        struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev);
        unsigned long max_apqi = matrix_mdev->matrix.aqm_max;
 
@@ -1008,6 +1081,10 @@ static ssize_t assign_domain_store(struct device *dev,
                goto share_err;
 
        vfio_ap_mdev_link_domain(matrix_mdev, apqi);
+       memset(aqm_delta, 0, sizeof(aqm_delta));
+       set_bit_inv(apqi, aqm_delta);
+       vfio_ap_mdev_filter_matrix(matrix_mdev->matrix.apm, aqm_delta,
+                                  matrix_mdev);
        ret = count;
        goto done;
 
@@ -1076,6 +1153,10 @@ static ssize_t unassign_domain_store(struct device *dev,
 
        clear_bit_inv((unsigned long)apqi, matrix_mdev->matrix.aqm);
        vfio_ap_mdev_unlink_domain(matrix_mdev, apqi);
+
+       if (test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm))
+               clear_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm);
+
        ret = count;
 
 done:
@@ -1129,6 +1210,7 @@ static ssize_t assign_control_domain_store(struct device *dev,
         * number of control domains that can be assigned.
         */
        set_bit_inv(id, matrix_mdev->matrix.adm);
+       vfio_ap_mdev_filter_cdoms(matrix_mdev);
        ret = count;
 done:
        mutex_unlock(&matrix_dev->lock);
@@ -1176,6 +1258,10 @@ static ssize_t unassign_control_domain_store(struct device *dev,
        }
 
        clear_bit_inv(domid, matrix_mdev->matrix.adm);
+
+       if (test_bit_inv(domid, matrix_mdev->shadow_apcb.adm))
+               clear_bit_inv(domid, matrix_mdev->shadow_apcb.adm);
+
        ret = count;
 done:
        mutex_unlock(&matrix_dev->lock);
@@ -1309,8 +1395,6 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
 
                kvm_get_kvm(kvm);
                matrix_mdev->kvm = kvm;
-               memcpy(&matrix_mdev->shadow_apcb, &matrix_mdev->matrix,
-                      sizeof(struct ap_matrix));
                kvm_arch_crypto_set_masks(kvm, matrix_mdev->shadow_apcb.apm,
                                          matrix_mdev->shadow_apcb.aqm,
                                          matrix_mdev->shadow_apcb.adm);
@@ -1687,6 +1771,11 @@ int vfio_ap_mdev_probe_queue(struct ap_device *apdev)
        q->apqn = to_ap_queue(&apdev->device)->qid;
        q->saved_isc = VFIO_AP_ISC_INVALID;
        vfio_ap_queue_link_mdev(q);
+       if (q->matrix_mdev) {
+               vfio_ap_mdev_filter_matrix(q->matrix_mdev->matrix.apm,
+                                          q->matrix_mdev->matrix.aqm,
+                                          q->matrix_mdev);
+       }
        dev_set_drvdata(&apdev->device, q);
        mutex_unlock(&matrix_dev->lock);
 
@@ -1695,15 +1784,21 @@ int vfio_ap_mdev_probe_queue(struct ap_device *apdev)
 
 void vfio_ap_mdev_remove_queue(struct ap_device *apdev)
 {
+       unsigned long apid;
        struct vfio_ap_queue *q;
 
        mutex_lock(&matrix_dev->lock);
        sysfs_remove_group(&apdev->device.kobj, &vfio_queue_attr_group);
        q = dev_get_drvdata(&apdev->device);
 
-       if (q->matrix_mdev)
+       if (q->matrix_mdev) {
                vfio_ap_unlink_queue_fr_mdev(q);
 
+               apid = AP_QID_CARD(q->apqn);
+               if (test_bit_inv(apid, q->matrix_mdev->shadow_apcb.apm))
+                       clear_bit_inv(apid, q->matrix_mdev->shadow_apcb.apm);
+       }
+
        vfio_ap_mdev_reset_queue(q, 1);
        dev_set_drvdata(&apdev->device, NULL);
        kfree(q);