virtio-vfc implementation

Implement the 'rescan' virtio-scsi feature. Rescanning works by sending a 'rescan' virtio-scsi command with the next requested target id to the backend. The backend will respond with the next used target id or '-1' if no more targets are found. This avoids scanning all possible targets.

Signed-off-by Hannes Reinecke <[email protected]>

---
  drivers/scsi/virtio_scsi.c       | 239 ++++++++++++++++++++++++++++++++++++++-
  include/uapi/linux/virtio_scsi.h |  15 +++
  2 files changed, 250 insertions(+), 4 deletions(-)

drivers/scsi/virtio_scsi.c +21 -0

index 7c28e8d..a561e90 100644
 
+static void virtscsi_rescan_work(struct work_struct *work)
+{
 
+	if (target_id == -1) {
+		shost_printk(KERN_INFO, sh, "rescan: terminated\n");
+		spin_unlock_irq(&vscsi->rescan_lock);
+		return;
+	}
+	spin_unlock_irq(&vscsi->rescan_lock);
+
+	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
+	if (!cmd) {
+		shost_printk(KERN_INFO, sh, "rescan: no memory\n");
+		goto scan_host;
+	}
+	shost_printk(KERN_INFO, sh, "rescan: next target %d\n", target_id);
 
+		shost_printk(KERN_INFO, sh,
+			     "rescan: no more targets\n");
 
+	shost_printk(KERN_INFO, sh, "rescan: scan host\n");
+	scsi_scan_host(sh);
+}
+

drivers/scsi/virtio_scsi.c +22 -0

+static void virtscsi_scan_start(struct Scsi_Host *sh)
+{
+	struct virtio_scsi *vscsi = shost_priv(sh);
+
+	virtscsi_scan_host(vscsi);
+	spin_lock_irq(&vscsi->rescan_lock);
+	if (vscsi->next_target_id != -1) {
+		shost_printk(KERN_INFO, sh, "rescan: already running\n");
+		spin_unlock_irq(&vscsi->rescan_lock);
+		return;
+	}
+	vscsi->next_target_id = 0;
+	shost_printk(KERN_INFO, sh, "rescan: start\n");
+	spin_unlock_irq(&vscsi->rescan_lock);
+	queue_work(system_freezable_wq, &vscsi->rescan_work);
+}
+
+int virtscsi_scan_finished(struct Scsi_Host *sh, unsigned long time)
 
+	shost_printk(KERN_INFO, sh, "rescan: %s finished\n",
+		     ret ? "" : "not");
+	return ret;
+}

When a device announces an 'FC' protocol we should be pulling in the FC transport class to have the rports etc setup correctly.

Signed-off-by Hannes Reinecke <[email protected]>

---
  drivers/scsi/virtio_scsi.c | 323 ++++++++++++++++++++++++++++++++++++++-------
  1 file changed, 277 insertions(+), 46 deletions(-)

drivers/scsi/virtio_scsi.c +10 -0

index a561e90..f925fbd 100644
@@ -591,11 +616,25 @@ static int virtscsi_queuecommand_single(struct Scsi_Host *sh,
  					struct scsi_cmnd *sc)
 
+	if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+		struct fc_rport *rport =
+			starget_to_rport(scsi_target(sc->device));
+		if (rport && rport->dd_data) {
+			tgt = rport->dd_data;
+			target_id = tgt->target_id;
+		}
+	} else
+		tgt = scsi_target(sc->device)->hostdata;
+	if (!tgt || tgt->removed) {
 

@@ -648,16 +687,30 @@ static int virtscsi_queuecommand_multi(struct Scsi_Host *sh,
  				       struct scsi_cmnd *sc)
 

drivers/scsi/virtio_scsi.c +5 -0

+			return FAST_IO_FAIL;
+	}
 
 

@@ -755,19 +823,34 @@ static int virtscsi_abort(struct scsi_cmnd *sc)
 
+	if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+		struct fc_rport *rport =
+			starget_to_rport(scsi_target(sc->device));

drivers/scsi/virtio_scsi.c +0 -6

  	target_id = virtio32_to_cpu(vscsi->vdev, cmd->resp.rescan.id);
-	if (target_id != -1) {
-		int transport = virtio32_to_cpu(vscsi->vdev,
-						cmd->resp.rescan.transport);
-		spin_lock_irq(&vscsi->rescan_lock);
-		vscsi->next_target_id = target_id + 1;
-		spin_unlock_irq(&vscsi->rescan_lock);

drivers/scsi/virtio_scsi.c +14 -1

+		} else {
+			spin_lock_irq(&vscsi->rescan_lock);
+			list_del(&tgt->list);
+			spin_unlock_irq(&vscsi->rescan_lock);
+			kfree(tgt);
+			tgt = NULL;
+		}
+	} else {
+		scsi_scan_target(&sh->shost_gendev, 0, tgt->target_id,
+				 SCAN_WILD_CARD, SCSI_SCAN_INITIAL);
+	}
+	queue_work(system_freezable_wq, &vscsi->rescan_work);
+out:
  	mempool_free(cmd, virtscsi_cmd_pool);
  	return;
  scan_host:
 

@@ -932,14 +1089,31 @@ static void virtscsi_scan_host(struct virtio_scsi *vscsi)
  static void virtscsi_scan_start(struct Scsi_Host *sh)
  {
  	struct virtio_scsi *vscsi = shost_priv(sh);
+	struct virtio_scsi_target_state *tgt;
 
-	virtscsi_scan_host(vscsi);
  	spin_lock_irq(&vscsi->rescan_lock);

drivers/scsi/virtio_scsi.c +8 -0

+				tgt->rport = NULL;
+			}
+			tgt->removed = true;
+		}
+	} else {
+		list_for_each_entry(tgt, &vscsi->target_list, list)
+			tgt->removed = true;
+	}
  	vscsi->next_target_id = 0;
  	shost_printk(KERN_INFO, sh, "rescan: start\n");
  	spin_unlock_irq(&vscsi->rescan_lock);

@@ -954,12 +1128,14 @@ int virtscsi_scan_finished(struct Scsi_Host *sh, unsigned long time)
  	spin_lock_irq(&vscsi->rescan_lock);
  	if (vscsi->next_target_id != -1)
  		ret = 0;

drivers/scsi/virtio_scsi.c +10 -0

  	return ret;
  }
 

@@ -978,6 +1154,34 @@ static ssize_t virtscsi_host_store_rescan(struct device *dev,
  	NULL,
  };
 
+static int virtscsi_issue_lip(struct Scsi_Host *shost)
+{
+	struct virtio_scsi *vscsi = shost_priv(shost);
+	unsigned long start = jiffies;
+	struct virtio_scsi_target_state *tgt;
+
+	spin_lock_irq(&vscsi->rescan_lock);
+	if (vscsi->next_target_id != -1) {
+		spin_unlock_irq(&vscsi->rescan_lock);
+		return 0;

drivers/scsi/virtio_scsi.c +9 -0

+	spin_unlock_irq(&vscsi->rescan_lock);
+	queue_work(system_freezable_wq, &vscsi->rescan_work);
+
+	while (!virtscsi_scan_finished(shost, jiffies - start))
+		msleep(10);
+
+	return 0;
+}
+
  static struct scsi_host_template virtscsi_host_template_single = {
  	.module = THIS_MODULE,

.name = "Virtio SCSI HBA",

drivers/scsi/virtio_scsi.c +14 -0

@@ -1066,6 +1270,20 @@ static ssize_t virtscsi_host_store_rescan(struct device *dev,
  	.track_queue_depth = 1,
  };
 
+static struct fc_function_template virtscsi_transport_functions = {
+	.dd_fcrport_size = sizeof(struct virtio_scsi_target_state *),
+	.show_host_node_name = 1,
+	.show_host_port_name = 1,
+	.show_host_port_id = 1,
+	.show_host_port_state = 1,
+	.show_host_port_type = 1,
+	.show_starget_node_name = 1,
+	.show_starget_port_name = 1,
+	.show_starget_port_id = 1,
+	.show_rport_dev_loss_tmo = 1,
+	.issue_fc_host_lip = virtscsi_issue_lip,
+};
+
  #define virtscsi_config_get(vdev, fld) \
  	({ \
  		typeof(((struct virtio_scsi_config *)0)->fld) __val; \

drivers/scsi/virtio_scsi.c +1 -0

  	spin_lock_init(&vscsi->rescan_lock);
+	INIT_LIST_HEAD(&vscsi->target_list);
  	INIT_WORK(&vscsi->rescan_work, virtscsi_rescan_work);
 
  	err = virtscsi_init(vdev, vscsi);

Quoted 2 lines · Hannes Reinecke

When a device announces an 'FC' protocol we should be pulling in the FC transport class to have the rports etc setup correctly.

It took some time for me to understand what this does.
It seems to mirror the topology of rports and sdevs that exist under the fc_host on the kvm host side to the virtio-scsi-fc side in the guest.

I like the idea. This is also what I've been suggesting users to do if they back virtio-scsi with zfcp on the kvm host side. Primarily to not stall all virtio-scsi I/O on all paths if the guest ever gets into scsi_eh. But also to make it look like an HBA pass through so one can more easily migrate to this once we have FCP pass through.

Quoted 31 lines

Signed-off-by Hannes Reinecke <[email protected]>

---
  drivers/scsi/virtio_scsi.c | 323
++++++++++++++++++++++++++++++++++++++-------

  1 file changed, 277 insertions(+), 46 deletions(-)

drivers/scsi/virtio_scsi.c +10 -0

index a561e90..f925fbd 100644
@@ -591,11 +616,25 @@ static int virtscsi_queuecommand_single(struct
 Scsi_Host *sh,
                       struct scsi_cmnd *sc)
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));
+        if (rport && rport->dd_data) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;
+        }
+    } else
+        tgt = scsi_target(sc->device)->hostdata;
+    if (!tgt || tgt->removed) {
 

@@ -648,16 +687,30 @@ static int virtscsi_queuecommand_multi(struct
 Scsi_Host *sh,
                          struct scsi_cmnd *sc)
 

drivers/scsi/virtio_scsi.c static int virtscsi_queuecommand_multi(struct · line 699

Quoted 10 lines

drivers/scsi/virtio_scsi.c +10 -0

+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));
+        if (rport && rport->dd_data) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;
+        }
+    } else
+        tgt = scsi_target(sc->device)->hostdata;
+    if (!tgt || tgt->removed) {

repeating pattern?

Quoted 9 lines

drivers/scsi/virtio_scsi.c +6 -0

@@ -696,12 +749,27 @@ static int virtscsi_device_reset(struct
 scsi_cmnd *sc)
   {
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));
+        if (rport && rport->dd_data ) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;

drivers/scsi/virtio_scsi.c static int virtscsi_device_reset(struct scsi_cmnd *sc) · line 770

Quoted 5 lines

drivers/scsi/virtio_scsi.c +5 -0

+        } else
+            return FAST_IO_FAIL;
+    } else {
+        tgt = scsi_target(sc->device)->hostdata;
+        if (!tgt || tgt->removed)

The other patterns have the !tgt check outside of the if else condition. For consistency this might work here, too?

Quoted 9 lines

drivers/scsi/virtio_scsi.c +5 -0

+            return FAST_IO_FAIL;
+    }
 
 

@@ -755,19 +823,34 @@ static int virtscsi_abort(struct scsi_cmnd *sc)
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));

drivers/scsi/virtio_scsi.c static int virtscsi_abort(struct scsi_cmnd *sc) · line 836

Quoted 10 lines

drivers/scsi/virtio_scsi.c +10 -0

+        if (rport && rport->dd_data ) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;
+        } else
+            return FAST_IO_FAIL;
+    } else {
+        tgt = scsi_target(sc->device)->hostdata;
+        if (!tgt || tgt->removed)
+            return FAST_IO_FAIL;
+    }

dito

Quoted 3 lines

drivers/scsi/virtio_scsi.c +0 -0

@@ -857,27 +970,67 @@ static void virtscsi_rescan_work(struct
 work_struct *work)
 

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 971

Quoted 1 line

drivers/scsi/virtio_scsi.c +0 -0

       wait_for_completion(&comp);

Waiting in work item .vs. having the response (IRQ) path trigger subsequent processing async ?
Or do we need the call chain(s) getting here to be in our own process context via the workqueue anyway?

Quoted 7 lines

drivers/scsi/virtio_scsi.c +0 -6

       target_id = virtio32_to_cpu(vscsi->vdev, cmd->resp.rescan.id);
-    if (target_id != -1) {
-        int transport = virtio32_to_cpu(vscsi->vdev,
-                        cmd->resp.rescan.transport);
-        spin_lock_irq(&vscsi->rescan_lock);
-        vscsi->next_target_id = target_id + 1;
-        spin_unlock_irq(&vscsi->rescan_lock);

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 973

Quoted 10 lines

drivers/scsi/virtio_scsi.c +1 -9

-        shost_printk(KERN_INFO, sh,
-                 "found %s target %d (WWN %*phN)\n",
-                 transport == SCSI_PROTOCOL_FCP ? "FC" : "SAS",
-                 target_id, 8,
-                 cmd->resp.rescan.port_wwn);
-        scsi_scan_target(&sh->shost_gendev, 0, target_id,
-                 SCAN_WILD_CARD, SCSI_SCAN_INITIAL);
-        queue_work(system_freezable_wq, &vscsi->rescan_work);
-    } else {
+    if (target_id == -1) {

This boolean if expression was introduced in the previous patch. Why negate it here?
It causes a number of potentially unnecessary changed lines making review harder.

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 1015

Quoted 8 lines

drivers/scsi/virtio_scsi.c +7 -0

+    if (transport == SCSI_PROTOCOL_FCP) {
+        struct fc_rport_identifiers rport_ids;
+        struct fc_rport *rport;
+
+        rport_ids.node_name = wwn_to_u64(cmd->resp.rescan.node_wwn);
+        rport_ids.port_name = wwn_to_u64(cmd->resp.rescan.port_wwn);
+        rport_ids.port_id = (target_id >> 8);

Why do you shift target_id by one byte to the right?

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 1020

Quoted 5 lines

drivers/scsi/virtio_scsi.c +5 -0

+        rport = fc_remote_port_add(sh, 0, &rport_ids);
+        if (rport) {
+            tgt->rport = rport;
+            rport->dd_data = tgt;
+            fc_remote_port_rolechg(rport, FC_RPORT_ROLE_FCP_TARGET);

Is the rolechg to get some event? Otherwise we could have rport_ids.roles = FC_RPORT_ROLE_FCP_TARGET before fc_remote_port_add().

Quoted 26 lines

drivers/scsi/virtio_scsi.c +13 -0

+        } else {
+            spin_lock_irq(&vscsi->rescan_lock);
+            list_del(&tgt->list);
+            spin_unlock_irq(&vscsi->rescan_lock);
+            kfree(tgt);
+            tgt = NULL;
+        }
+    } else {
+        scsi_scan_target(&sh->shost_gendev, 0, tgt->target_id,
+                 SCAN_WILD_CARD, SCSI_SCAN_INITIAL);
+    }
+    queue_work(system_freezable_wq, &vscsi->rescan_work);
+out:
       mempool_free(cmd, virtscsi_cmd_pool);
       return;

  scan_host:

drivers/scsi/virtio_scsi.c +1 -1

@@ -932,14 +1089,31 @@ static void virtscsi_scan_host(struct
 virtio_scsi *vscsi)
   static void virtscsi_scan_start(struct Scsi_Host *sh)
   {
       struct virtio_scsi *vscsi = shost_priv(sh);
+    struct virtio_scsi_target_state *tgt;
 
-    virtscsi_scan_host(vscsi);
       spin_lock_irq(&vscsi->rescan_lock);

drivers/scsi/virtio_scsi.c static void virtscsi_scan_host(struct · line 1105

Quoted 10 lines

drivers/scsi/virtio_scsi.c +5 -0

       if (vscsi->next_target_id != -1) {
           shost_printk(KERN_INFO, sh, "rescan: already running\n");
           spin_unlock_irq(&vscsi->rescan_lock);
           return;
       }
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        fc_host_node_name(sh) = vscsi->wwnn;
+        fc_host_port_name(sh) = vscsi->wwpn;
+        fc_host_port_id(sh) = 0x00ff00;
+        fc_host_port_type(sh) = FC_PORTTYPE_NPIV;

Why is this hardcoded?

At least with zfcp, we can have kvm host *v*HBAs without NPIV.

drivers/scsi/virtio_scsi.c static void virtscsi_scan_host(struct virtio_scsi *vscsi) · line 1108

Quoted 4 lines

drivers/scsi/virtio_scsi.c +4 -0

+        fc_host_port_state(sh) = FC_PORTSTATE_BLOCKED;
+        list_for_each_entry(tgt, &vscsi->target_list, list) {
+            if (tgt->rport) {
+                fc_remote_port_delete(tgt->rport);

I'm not familiar with the scan callbacks, maybe that's why I wonder why you block all rports when scanning the host. Is it so that the subsequent port scans can call the properly paired fc_remote_port_add to potentially update a previously existing rport with new properties we got from the qemu/host side.

Quoted 22 lines

drivers/scsi/virtio_scsi.c +8 -0

+                tgt->rport = NULL;
+            }
+            tgt->removed = true;
+        }
+    } else {
+        list_for_each_entry(tgt, &vscsi->target_list, list)
+            tgt->removed = true;
+    }
       vscsi->next_target_id = 0;
       shost_printk(KERN_INFO, sh, "rescan: start\n");

      spin_unlock_irq(&vscsi->rescan_lock);

drivers/scsi/virtio_scsi.c +2 -0

@@ -954,12 +1128,14 @@ int virtscsi_scan_finished(struct Scsi_Host
 *sh, unsigned long time)
       spin_lock_irq(&vscsi->rescan_lock);
       if (vscsi->next_target_id != -1)
           ret = 0;
+    else if (vscsi->protocol == SCSI_PROTOCOL_FCP)
+        fc_host_port_state(sh) = FC_PORTSTATE_ONLINE;
       spin_unlock_irq(&vscsi->rescan_lock);
       if (!ret)
           flush_work(&vscsi->rescan_work);
 

drivers/scsi/virtio_scsi.c int virtscsi_scan_finished(struct Scsi_Host *sh, unsigned long time) · line 1138

Quoted 4 lines

drivers/scsi/virtio_scsi.c +2 -2

-    shost_printk(KERN_INFO, sh, "rescan: %s finished\n",
-             ret ? "" : "not");
+    shost_printk(KERN_INFO, sh, "rescan: %sfinished\n",
+             ret ? "" : "not ");

You introduced it in the previous patch, so do the fixup there in the first place?

Quoted 18 lines

drivers/scsi/virtio_scsi.c +10 -0

       return ret;
   }

@@ -978,6 +1154,34 @@ static ssize_t virtscsi_host_store_rescan(struct
 device *dev,
       NULL,
   };
 
+static int virtscsi_issue_lip(struct Scsi_Host *shost)
+{
+    struct virtio_scsi *vscsi = shost_priv(shost);
+    unsigned long start = jiffies;
+    struct virtio_scsi_target_state *tgt;
+
+    spin_lock_irq(&vscsi->rescan_lock);
+    if (vscsi->next_target_id != -1) {
+        spin_unlock_irq(&vscsi->rescan_lock);
+        return 0;

drivers/scsi/virtio_scsi.c static ssize_t virtscsi_host_store_rescan(struct · line 1177

Quoted 10 lines

drivers/scsi/virtio_scsi.c +9 -0

+    }
 
+    fc_host_port_state(shost) = FC_PORTSTATE_BLOCKED;
+    list_for_each_entry(tgt, &vscsi->target_list, list) {
+        if (tgt->rport) {
+            fc_remote_port_delete(tgt->rport);
+            tgt->rport = NULL;
+        }
+    }
+    vscsi->next_target_id = 0;

I see some code duplication with what's in virtscsi_scan_host(). Not sure if reduction is worth it.

Quoted 37 lines

drivers/scsi/virtio_scsi.c +9 -0

+    spin_unlock_irq(&vscsi->rescan_lock);
+    queue_work(system_freezable_wq, &vscsi->rescan_work);
+
+    while (!virtscsi_scan_finished(shost, jiffies - start))
+        msleep(10);
+
+    return 0;
+}
+
   static struct scsi_host_template virtscsi_host_template_single = {

      .module = THIS_MODULE,
      .name = "Virtio SCSI HBA",

drivers/scsi/virtio_scsi.c +14 -0

@@ -1066,6 +1270,20 @@ static ssize_t
 virtscsi_host_store_rescan(struct device *dev,
       .track_queue_depth = 1,
   };
 
+static struct fc_function_template virtscsi_transport_functions = {
+    .dd_fcrport_size = sizeof(struct virtio_scsi_target_state *),
+    .show_host_node_name = 1,
+    .show_host_port_name = 1,
+    .show_host_port_id = 1,
+    .show_host_port_state = 1,
+    .show_host_port_type = 1,
+    .show_starget_node_name = 1,
+    .show_starget_port_name = 1,
+    .show_starget_port_id = 1,
+    .show_rport_dev_loss_tmo = 1,
+    .issue_fc_host_lip = virtscsi_issue_lip,
+};
+
   #define virtscsi_config_get(vdev, fld) \
       ({ \

          typeof(((struct virtio_scsi_config *)0)->fld) __val; \

drivers/scsi/virtio_scsi.c +0 -0

@@ -1193,7 +1411,9 @@ static int virtscsi_probe(struct virtio_device
 *vdev)
       vscsi->num_queues = num_queues;

drivers/scsi/virtio_scsi.c static int virtscsi_probe(struct virtio_device *vdev) · line 1414

Quoted 3 lines

drivers/scsi/virtio_scsi.c +1 -0

       vdev->priv = shost;
       vscsi->next_target_id = -1;
+    vscsi->protocol = SCSI_PROTOCOL_SAS;

Why is the old/legacy/non-fcp hardcoded SAS?
Doesn't the non-fcp virtio-scsi have any real transport at all, i.e. "none"?
Maybe I just don't understand semantics of vscsi->protocol well enough.

Quoted 8 lines · Steffen Maier

Quoted 2 lines · Hannes Reinecke

When a device announces an 'FC' protocol we should be pulling in the FC transport class to have the rports etc setup correctly.

It took some time for me to understand what this does.
It seems to mirror the topology of rports and sdevs that exist under the fc_host on the kvm host side to the virtio-scsi-fc side in the guest.

Quoted 5 lines

I like the idea. This is also what I've been suggesting users to do if they back virtio-scsi with zfcp on the kvm host side. Primarily to not stall all virtio-scsi I/O on all paths if the guest ever gets into scsi_eh. But also to make it look like an HBA pass through so one can more easily migrate to this once we have FCP pass through.

Quoted 21 lines

Quoted 5 lines

drivers/scsi/virtio_scsi.c +3 -0

@@ -755,19 +823,34 @@ static int virtscsi_abort(struct scsi_cmnd *sc)
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));

drivers/scsi/virtio_scsi.c static int virtscsi_abort(struct scsi_cmnd *sc) · line 836

Quoted 10 lines

drivers/scsi/virtio_scsi.c +10 -0

+        if (rport && rport->dd_data ) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;
+        } else
+            return FAST_IO_FAIL;
+    } else {
+        tgt = scsi_target(sc->device)->hostdata;
+        if (!tgt || tgt->removed)
+            return FAST_IO_FAIL;
+    }

dito

Quoted 3 lines

(quoted diff) +0 -0

@@ -857,27 +970,67 @@ static void virtscsi_rescan_work(struct
 work_struct *work)
 

Quoted 7 lines

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 971

Quoted 1 line

drivers/scsi/virtio_scsi.c +0 -0

       wait_for_completion(&comp);

Waiting in work item .vs. having the response (IRQ) path trigger subsequent processing async ?
Or do we need the call chain(s) getting here to be in our own process context via the workqueue anyway?

Can't see I can parse this sentence, but I'll be looking at the code trying to come up with a clever explanation :-)

Quoted 8 lines

Quoted 7 lines

drivers/scsi/virtio_scsi.c +7 -0

+    if (transport == SCSI_PROTOCOL_FCP) {
+        struct fc_rport_identifiers rport_ids;
+        struct fc_rport *rport;
+
+        rport_ids.node_name = wwn_to_u64(cmd->resp.rescan.node_wwn);
+        rport_ids.port_name = wwn_to_u64(cmd->resp.rescan.port_wwn);
+        rport_ids.port_id = (target_id >> 8);

Quoted 2 lines

Why do you shift target_id by one byte to the right?

Because with the original setup virtio_scsi guest would pass in the target_id, and the host would be selecting the device based on that information.
With virtio-vfc we pass in the wwpn, but still require the target ID to be compliant with things like event notification etc.

So I've shifted the target id onto the port ID (which is 24 bit anyway). I could've used a bitfield here, but then I wasn't quite sure about the endianness of which.

Quoted 6 lines

Quoted 5 lines

drivers/scsi/virtio_scsi.c +5 -0

+        rport = fc_remote_port_add(sh, 0, &rport_ids);
+        if (rport) {
+            tgt->rport = rport;
+            rport->dd_data = tgt;
+            fc_remote_port_rolechg(rport, FC_RPORT_ROLE_FCP_TARGET);

Quoted 3 lines

Is the rolechg to get some event? Otherwise we could have rport_ids.roles = FC_RPORT_ROLE_FCP_TARGET before fc_remote_port_add().

That's how the 'normal' transport classes do it; but I'll check if this can be rolled into the call to fc_remote_port_add().

Quoted 11 lines

Quoted 10 lines

drivers/scsi/virtio_scsi.c +5 -0

@@ -932,14 +1089,31 @@ static void virtscsi_scan_host(struct
 virtio_scsi *vscsi)
   static void virtscsi_scan_start(struct Scsi_Host *sh)
   {
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        fc_host_node_name(sh) = vscsi->wwnn;
+        fc_host_port_name(sh) = vscsi->wwpn;
+        fc_host_port_id(sh) = 0x00ff00;
+        fc_host_port_type(sh) = FC_PORTTYPE_NPIV;

Quoted 4 lines

Why is this hardcoded?

At least with zfcp, we can have kvm host *v*HBAs without NPIV.

For the simple fact that I don't have fields to transfer this information; plus it's not _actually_ used anywhere.

Unless you have a strict need why this information is required?

Quoted 9 lines · Hannes Reinecke

Quoted 8 lines · Steffen Maier

Quoted 2 lines · Hannes Reinecke

When a device announces an 'FC' protocol we should be pulling in the FC transport class to have the rports etc setup correctly.

It took some time for me to understand what this does.
It seems to mirror the topology of rports and sdevs that exist under the fc_host on the kvm host side to the virtio-scsi-fc side in the guest.

Quoted 5 lines

Quoted 5 lines

I like the idea. This is also what I've been suggesting users to do if they back virtio-scsi with zfcp on the kvm host side. Primarily to not stall all virtio-scsi I/O on all paths if the guest ever gets into scsi_eh. But also to make it look like an HBA pass through so one can more easily migrate to this once we have FCP pass through.

On second thought, I like the idea for virtio-scsi.

For the future virtio-(v)fc case, see below.

Quoted 28 lines

Quoted 21 lines

Quoted 5 lines

drivers/scsi/virtio_scsi.c +3 -0

@@ -755,19 +823,34 @@ static int virtscsi_abort(struct scsi_cmnd *sc)
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        struct fc_rport *rport =
+            starget_to_rport(scsi_target(sc->device));

drivers/scsi/virtio_scsi.c static int virtscsi_abort(struct scsi_cmnd *sc) · line 836

Quoted 10 lines

drivers/scsi/virtio_scsi.c +10 -0

+        if (rport && rport->dd_data ) {
+            tgt = rport->dd_data;
+            target_id = tgt->target_id;
+        } else
+            return FAST_IO_FAIL;
+    } else {
+        tgt = scsi_target(sc->device)->hostdata;
+        if (!tgt || tgt->removed)
+            return FAST_IO_FAIL;
+    }

dito

Quoted 3 lines

(quoted diff) +0 -0

@@ -857,27 +970,67 @@ static void virtscsi_rescan_work(struct
 work_struct *work)
 

Quoted 7 lines

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 971

Quoted 1 line

drivers/scsi/virtio_scsi.c +0 -0

        wait_for_completion(&comp);

Waiting in work item .vs. having the response (IRQ) path trigger subsequent processing async ?
Or do we need the call chain(s) getting here to be in our own process context via the workqueue anyway?

Quoted 2 lines

Can't see I can parse this sentence, but I'll be looking at the code trying to come up with a clever explanation :-)

Sorry, meanwhile I have a hard time understanding my own words, too.

I think I wondered if the effort of a work item is really necessary, especially considering that it does block on the completion and thus could delay other queued work items (even though Concurrency Managed Workqueues can often hide this delay).

Couldn't we just return asynchronously after having sent the request. And then later on, simply have the response (IRQ) path trigger whatever processing is necessary (after the work item variant woke up from the wait_for_completion) in some asynchronuous fashion? Of course, this could also be a work item which just does necessary remaining processing after we got a response.
Just a wild guess, without knowing the environmental requirements.

Quoted 8 lines

Quoted 8 lines

Quoted 7 lines

drivers/scsi/virtio_scsi.c +7 -0

+    if (transport == SCSI_PROTOCOL_FCP) {
+        struct fc_rport_identifiers rport_ids;
+        struct fc_rport *rport;
+
+        rport_ids.node_name = wwn_to_u64(cmd->resp.rescan.node_wwn);
+        rport_ids.port_name = wwn_to_u64(cmd->resp.rescan.port_wwn);
+        rport_ids.port_id = (target_id >> 8);

Quoted 7 lines

Quoted 2 lines

Why do you shift target_id by one byte to the right?

Because with the original setup virtio_scsi guest would pass in the target_id, and the host would be selecting the device based on that information.
With virtio-vfc we pass in the wwpn, but still require the target ID to be compliant with things like event notification etc.

Don't we need the true N_Port-ID, then? That's what an fc_rport.port_id usually contains. It's also a simple way to lookup resources on a SAN switch for problem determination. Or did I misunderstand the content/semantics of the variable target_id, assuming it's a SCSI target ID, i.e. the 3rd part of a HCTL 4-tuple?

Quoted 13 lines

So I've shifted the target id onto the port ID (which is 24 bit anyway). I could've used a bitfield here, but then I wasn't quite sure about the endianness of which.

Quoted 9 lines

drivers/scsi/virtio_scsi.c static void virtscsi_rescan_work(struct work_struct *work) · line 1020

Quoted 5 lines

drivers/scsi/virtio_scsi.c +5 -0

+        rport = fc_remote_port_add(sh, 0, &rport_ids);
+        if (rport) {
+            tgt->rport = rport;
+            rport->dd_data = tgt;
+            fc_remote_port_rolechg(rport, FC_RPORT_ROLE_FCP_TARGET);

Is the rolechg to get some event? Otherwise we could have rport_ids.roles = FC_RPORT_ROLE_FCP_TARGET before fc_remote_port_add().

Quoted 2 lines

That's how the 'normal' transport classes do it; but I'll check if this can be rolled into the call to fc_remote_port_add().

My idea was just based on how zfcp does it. Do you think I need to check if zfcp should do it via rolechg (even though zfcp never changes an rport role since it can only open targets)?

Quoted 11 lines

Quoted 11 lines

Quoted 10 lines

drivers/scsi/virtio_scsi.c +5 -0

@@ -932,14 +1089,31 @@ static void virtscsi_scan_host(struct
 virtio_scsi *vscsi)
    static void virtscsi_scan_start(struct Scsi_Host *sh)
    {
 
+    if (vscsi->protocol == SCSI_PROTOCOL_FCP) {
+        fc_host_node_name(sh) = vscsi->wwnn;
+        fc_host_port_name(sh) = vscsi->wwpn;
+        fc_host_port_id(sh) = 0x00ff00;
+        fc_host_port_type(sh) = FC_PORTTYPE_NPIV;

Quoted 6 lines

Quoted 4 lines

Why is this hardcoded?

At least with zfcp, we can have kvm host *v*HBAs without NPIV.

For the simple fact that I don't have fields to transfer this information; plus it's not _actually_ used anywhere.

It might not be used in this code, but it is exported to user space via sysfs. If I understood things correctly, you newly introduce virtio-scsi commands in patch 1 and are free to add more fields to struct virtio_scsi_rescan_resp (maybe within some max size limit)? To me, this raises the question which properties of the host's FC (driver core) objects should be mirrored to the guest. Ideally all (and that's a lot).
This in turn makes me wonder if mirroring is really desirable (e.g. considering the effort) or if only the guest should have its own FC object hierarchy which does _not_ exist on the KVM host in case an fc_host is passed through with virtio-(v)fc.

Quoted 1 line

Unless you have a strict need why this information is required?

In general, I would wish that virtio-(v)fc has no hard requirement for creating real vports on the KVM host. From a zfcp perspective, it would be nice if already existing fc_hosts of any type (vport or not) could be configured for virtio-vfc pass through. A selection key could e.g. be fc_host.port_name which is the (NPIV) WWPN of the initiator.

Rationale:

A 1st level hypervisor manages physical devices.
A 1st level hypervisor can pass through devices to a 1st level guest.

On x86, the 1st level hypervisor is KVM.
It can pass through PCI devices (physical or virtual functions). It can create vports (defining the NPIV WWPN) to pass through with virtio-vfc.
From your presentation [1] I derived and as quite surprised that there does not seem to be a thing that combines a VF and vport (actually no VFs for HBAs at all), because then x86 would already have everything
needed: Just PCI pass through a vport-VF and be done.
If I understood your slides correctly "mediated device passthrough" would be a software "workaround" because the hardware does not provide vport-VFs.

(The following assumes an FCP channel / zfcp perspective.)
On s390, the 1st level hypervisor is PR/SM providing LPARs. The PR/SM ecosystem defines virtual function equivalents [2, slide 15]. They can either use NPIV (kind of a vport) or not (FCP channel hardware virtualization was there before NPIV). Admittedly, the latter case is probably of no(t so much) importance for a virtio-vfc use case {it would be more like virtio-fc, i.e. without the 'v' as in NPIV}.
For the NPIV case, we have something like a vport and VF combined into a device, that we can sense on the system bus CCW (would be PCI on x86). The PR/SM ecosystem also defines the virtual NPIV WWPN for such initiator. An LPAR only sees an equivalent of virtual functions; it does not get access to a PF equivalent.
Hence, zfcp can hardly reasonably implement the vport creation interface of scsi_transport_fc; it only fakes the fc_host.port_type but otherwise all zfcp fc_hosts are "real", i.e. non-vport.
KVM as a 2nd level hypervisor sees the VF equivalents of the LPAR it runs in. Hence, KVM can only take whatever devices are there, HBA LLDDs in that KVM cannot create vports themselves.
(BTW, in case of z/VM as 2nd level hypervisor, it supports live guest migration with zfcp devices passed through the 1st and the 2nd level hypervisor.)

IOW, it would be nice if virtio-vfc could support nested KVM as 2nd level hypervisor just seeing already existing VFs or vports, which are in turn managed, defined and passed through by some 1st level hypervisor.

From the guest side I just have exposed the port_id, wwpn, and wwnn, as

A few more thoughts on your presentation [1]:

"Devices on the vport will not be visible on the host"
I could not agree more to the design point that devices (or at least their descendant object subtree) passed through to a guest should not appear on the host!
With virtio-blk or virtio-scsi, we have SCSI devices and thus disks visible in the host, which needlessly scans partitions, or even worse automatically scans for LVM and maybe even activates PVs/VGs/LVs. It's hard for a KVM host admin to suppress this (and not break the devices the host needs itself).
If we mirror the host's scsi_transport_fc tree including fc_rports and

thus SCSI devices etc., we would have the same problems? Even more so, dev_loss_tmo and fast_io_fail_tmo would run independently on the host and in the guest on the same mirrored scsi_transport_fc object tree. I can envision user confusion having configured timeouts on the "wrong" side (host vs. guest). Also we would still need a mechanism to mirror fc_rport (un)block from host to guest for proper transport recovery. In zfcp we try to recover on transport rather than scsi_eh whenever possible because it is so much smoother.

"Towards virtio-fc?"
Using the FCP_CMND_IU (instead of just a plain SCB as with virtio-scsi) sounds promising to me as starting point.
A listener from the audience asked if you would also do ELS/CT in the guest and you replied that this would not be good. Why is that? Based on above starting point, doing ELS/CT (and basic aborts and maybe a few other functions such as open/close ports or metadata transfer commands) in the guest is exactly what I would have expected. An HBA

LLDD on the KVM host would implement such API and for all fc_hosts, passed through this way, it would *not* establish any scsi_transport_fc tree on the host. Instead the one virtio-vfc implementation in the guest would do this indendently of which HBA LLDD provides the passed through fc_host in the KVM host.
ELS/CT pass through is maybe even for free via FC_BSG for those LLDDs that already implement it.
Rport open/close is just the analogon of slave_alloc()/slave_destroy().

A new tricky part would be how to "unbind" an fc_host on the KVM host to be able to pass it through. (At least on x86) we have a vport and thus would not have a system bus (PCI) device we could unbind and then bind to the virtio-vfc equivalent of VFIO for pass through. A virtio-vfc API with sysfs interface for HBA LLDDs on the host could provide something similar on an fc_host level.
Even better would be if an HBA LLDD would already know at probe time which of its fc_hosts are planned for pass through so it would not even

start establishing the whole FC transport object tree with port discovery and the implied SCSI target scan. A late unbind would, at least with zfcp, cause the object tree to linger with fc_rports transitioning to "not present" state eventually (because zfcp uses the default FC_TGTID_BIND_BY_WWPN and the fc_host lives long until module unload or system shutdown, so fc_remote_port_delete() does not really delete rports from sysfs) and SCSI devices transitioning to "transport-offline" eventually.

Quoted 24 lines · Steffen Maier

Quoted 2 lines · Hannes Reinecke

Quoted 1 line · Steffen Maier

On 12/14/2017 11:11 AM, Hannes Reinecke wrote:

To me, this raises the question which properties of the host's FC (driver core) objects should be mirrored to the guest. Ideally all (and that's a lot).
This in turn makes me wonder if mirroring is really desirable (e.g. considering the effort) or if only the guest should have its own FC object hierarchy which does _not_ exist on the KVM host in case an fc_host is passed through with virtio-(v)fc.

A few more thoughts on your presentation [1]:

"Devices on the vport will not be visible on the host"
I could not agree more to the design point that devices (or at least their descendant object subtree) passed through to a guest should not appear on the host!
With virtio-blk or virtio-scsi, we have SCSI devices and thus disks visible in the host, which needlessly scans partitions, or even worse automatically scans for LVM and maybe even activates PVs/VGs/LVs. It's hard for a KVM host admin to suppress this (and not break the devices the host needs itself).
If we mirror the host's scsi_transport_fc tree including fc_rports and

Quoted 9 lines

thus SCSI devices etc., we would have the same problems? Even more so, dev_loss_tmo and fast_io_fail_tmo would run independently on the host and in the guest on the same mirrored scsi_transport_fc object tree. I can envision user confusion having configured timeouts on the "wrong" side (host vs. guest). Also we would still need a mechanism to mirror fc_rport (un)block from host to guest for proper transport recovery. In zfcp we try to recover on transport rather than scsi_eh whenever possible because it is so much smoother.

As similar thing can be achieved event today, by setting the 'no_uld_attach' parameter when scanning the scsi device (that's what some RAID HBAs do).
However, there currently is no way of modifying it from user-space, and certainly not to change the behaviour for existing devices. It should be relatively simple to set this flag whenever the host is exposed to a VM; we would still see the scsi devices, but the 'sd' driver won't be attached so nothing will scan the device on the host.

Quoted 8 lines

"Towards virtio-fc?"
Using the FCP_CMND_IU (instead of just a plain SCB as with virtio-scsi) sounds promising to me as starting point.
A listener from the audience asked if you would also do ELS/CT in the guest and you replied that this would not be good. Why is that? Based on above starting point, doing ELS/CT (and basic aborts and maybe a few other functions such as open/close ports or metadata transfer commands) in the guest is exactly what I would have expected. An HBA

Quoted 9 lines

LLDD on the KVM host would implement such API and for all fc_hosts, passed through this way, it would *not* establish any scsi_transport_fc tree on the host. Instead the one virtio-vfc implementation in the guest would do this indendently of which HBA LLDD provides the passed through fc_host in the KVM host.
ELS/CT pass through is maybe even for free via FC_BSG for those LLDDs that already implement it.
Rport open/close is just the analogon of slave_alloc()/slave_destroy().

I'm not convinced that moving to full virtio-fc is something we want or even can do.
Neither qla2xxx nor lpfc allow for direct FC frame access; so one would need to reformat the FC frames into something the driver understands, just so that the hardware can transform it back into FC frames.

Another thing is xid management; some drivers have to do their own xid management, based on hardware capabilities etc.
So the FC frames would need to re-write the xids, making it hard if not impossible to match things up when the response comes in.

And, more importantly, what's the gain here?
Which feature do we miss?

Quoted 34 lines · Hannes Reinecke

Quoted 24 lines · Steffen Maier

Quoted 2 lines · Hannes Reinecke

Quoted 1 line · Steffen Maier

On 12/14/2017 11:11 AM, Hannes Reinecke wrote:

To me, this raises the question which properties of the host's FC (driver core) objects should be mirrored to the guest. Ideally all (and that's a lot).
This in turn makes me wonder if mirroring is really desirable (e.g. considering the effort) or if only the guest should have its own FC object hierarchy which does _not_ exist on the KVM host in case an fc_host is passed through with virtio-(v)fc.

A few more thoughts on your presentation [1]:

"Devices on the vport will not be visible on the host"
I could not agree more to the design point that devices (or at least their descendant object subtree) passed through to a guest should not appear on the host!
With virtio-blk or virtio-scsi, we have SCSI devices and thus disks visible in the host, which needlessly scans partitions, or even worse automatically scans for LVM and maybe even activates PVs/VGs/LVs. It's hard for a KVM host admin to suppress this (and not break the devices the host needs itself).
If we mirror the host's scsi_transport_fc tree including fc_rports and

Quoted 9 lines

thus SCSI devices etc., we would have the same problems? Even more so, dev_loss_tmo and fast_io_fail_tmo would run independently on the host and in the guest on the same mirrored scsi_transport_fc object tree. I can envision user confusion having configured timeouts on the "wrong" side (host vs. guest). Also we would still need a mechanism to mirror fc_rport (un)block from host to guest for proper transport recovery. In zfcp we try to recover on transport rather than scsi_eh whenever possible because it is so much smoother.

Quoted 8 lines

As similar thing can be achieved event today, by setting the 'no_uld_attach' parameter when scanning the scsi device (that's what some RAID HBAs do).
However, there currently is no way of modifying it from user-space, and certainly not to change the behaviour for existing devices. It should be relatively simple to set this flag whenever the host is exposed to a VM; we would still see the scsi devices, but the 'sd' driver won't be attached so nothing will scan the device on the host.

Ah, nice, didn't know that. It would solve the undesired I/O problem in the host.
But it would not solve the so far somewhat unsynchronized state transitions of fc_rports on the host and their mirrors in the guest?

I would be very interested in how you intend to do transport recovery.

Quoted 1 line

Quoted 1 line

"Towards virtio-fc?"

Quoted 5 lines

I'm not convinced that moving to full virtio-fc is something we want or even can do.
Neither qla2xxx nor lpfc allow for direct FC frame access; so one would need to reformat the FC frames into something the driver understands, just so that the hardware can transform it back into FC frames.

I thought of a more high level para-virtualized FCP HBA interface, than FC frames (which did exist in kernel v2.4 under drivers/fc4/ but no longer as it seems). Just like large parts of today's FCP LLDDs handle scatter gather lists and framing is done by the hardware.

Quoted 4 lines

Another thing is xid management; some drivers have to do their own xid management, based on hardware capabilities etc.
So the FC frames would need to re-write the xids, making it hard if not impossible to match things up when the response comes in.

For such things, where the hardware exposes more details (than, say, zfcp sees) I thought the LLDD on the KVM host would handle such details internally and only expose the higher level interface to virtio-fc.

Maybe something roughly like the basic transport protocol part of ibmvfc/ibmvscsi (not the other end in the firmware and not the cross partition DMA part), if I understood its overall design correctly by quickly looking at the code.
I somewhat had the impression that zfcp isn't too far from the overall operations style. As seem qla2xxx or lpfc to me, they just see and need to handle some more low-level FC details.
Conceptually replace CRQ/RDMA or QDIO with virtio.
(And for ibmvscsi also: SRP => FCP because it uses a different SCSI transport.)

Quoted 2 lines

And, more importantly, what's the gain here?
Which feature do we miss?

Pros: Full transport_fc tree in guest (only) and reliable fc_rport state transitions done in the guest, where I would expect them as user for a passed through fc_host. No effort needed for transport tree mirroring and trying to get (mirrored) rport state transitions right.

Cons: Of course such virtio-fc protocol/API would be a bit more than just FCP_CMND_IU and FCP_RSP_IU. It's also: abort, open/close port, open/close LUN, ELS, CT, get HBA metadata, and a mechanism for unsolicited notifications from the HBA LLDD on the KVM host side mainly for link up/down. And any host LLDD interested in supporting it would need some modifications to implement such API.

Admittedly, with your currently limited transport tree property mirroring it is likely less effort to get a first working virtio-vfc. However, I would not call it fc_host pass through. To me it currently more looks like a bit of a "faked" (no offense) transport tree in the guest and the rest is close to today's virtio-scsi?
I'm just trying to understand where this would be going, in order to get a conceptual impression or classification in my brain.

The 'native LUN' feature allows virtio-scsi to pass in the LUN numbers from the underlying storage directly, without having to modify the LUN number itself.
It works by shifting the existing LUN number down by 8 bytes, and add the virtio-specific 8-byte LUN steering header. With that virtio doesn't have to mangle the LUN number, allowing us to pass the 'real' LUN number to the guest.

Of course, we do cut off the last 8 bytes of the 'real' LUN number, but I'm not aware of any array utilizing that, so the impact should be negligible.

Signed-off-by Hannes Reinecke <[email protected]>

---
  drivers/scsi/virtio_scsi.c       | 62 ++++++++++++++++++++++++++++++----------
  include/uapi/linux/virtio_scsi.h |  1 +
  2 files changed, 48 insertions(+), 15 deletions(-)

drivers/scsi/virtio_scsi.c +5 -5

index f925fbd..63c2c85 100644
@@ -356,8 +356,12 @@ static void virtscsi_handle_transport_reset(struct virtio_scsi *vscsi,
  	struct scsi_device *sdev;
  	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
  	unsigned int target = event->lun[1];
-	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
+	u64 lun;
 
+	if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_NATIVE_LUN))
+		lun = scsilun_to_int((struct scsi_lun *)event->lun) >> 16;
+	else
+		lun = (event->lun[2] << 8) | event->lun[3];
 
 

@@ -524,10 +532,16 @@ static void virtio_scsi_init_hdr(struct virtio_device *vdev,
  				 int target_id,
  				 struct scsi_cmnd *sc)
  {
-	cmd->lun[0] = 1;
-	cmd->lun[1] = target_id;
-	cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
-	cmd->lun[3] = sc->device->lun & 0xff;

drivers/scsi/virtio_scsi.c +10 -2

-		.lun[2] = (sc->device->lun >> 8) | 0x40,
-		.lun[3] = sc->device->lun & 0xff,
  		.tag = cpu_to_virtio64(vscsi->vdev, (unsigned long)sc),
  	};
+	if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+		u64 lun = sc->device->lun << 16;
+		lun |= ((u64)1 << 8) | (u64)target_id;
+		int_to_scsilun(lun, (struct scsi_lun *)&cmd->req.tmf.lun);
+	} else {
+		cmd->req.tmf.lun[0] = 1;
+		cmd->req.tmf.lun[1] = target_id;
+		cmd->req.tmf.lun[2] = (sc->device->lun >> 8) | 0x40;
+		cmd->req.tmf.lun[3] = sc->device->lun & 0xff;
+	}
 
  	return virtscsi_tmf(vscsi, cmd);

  }

drivers/scsi/virtio_scsi.c +4 -1

-	shost->max_lun = virtscsi_config_get(vdev, max_lun) + 1 + 0x4000;
+	if (!virtio_has_feature(vdev, VIRTIO_SCSI_F_NATIVE_LUN))
+		shost->max_lun = virtscsi_config_get(vdev, max_lun) + 1 + 0x4000;
+	else
+		shost->max_lun = (u64)-1;
  	shost->max_id = num_targets;
  	shost->max_channel = 0;
  	shost->max_cmd_len = VIRTIO_SCSI_CDB_SIZE;

Just a few very early view-only review comments.
Haven't run the code.

Quoted 7 lines · Hannes Reinecke

The 'native LUN' feature allows virtio-scsi to pass in the LUN numbers from the underlying storage directly, without having to modify the LUN number itself.
It works by shifting the existing LUN number down by 8 bytes, and add the virtio-specific 8-byte LUN steering header. With that virtio doesn't have to mangle the LUN number, allowing us to pass the 'real' LUN number to the guest.

I only see shifts by 16 bits in the code below which would be 2 bytes. I had a quick look at the corresponding qemu code which looked the same to me.
What's the relation to 8 byte shifting, which would be 64 bit shift and thus odd for a 64 bit LUN, mentioned in the description here?

If the code keeps the LUN level 1 and 2 (dropping level 3 and 4) and I just don't understand it, it would be fine, I guess.

Quoted 3 lines

Of course, we do cut off the last 8 bytes of the 'real' LUN number, but I'm not aware of any array utilizing that, so the impact should be negligible.

Why did we do v3.17 commit 9cb78c16f5da ("scsi: use 64-bit LUNs")? ;-)

Quoted 40 lines

Signed-off-by Hannes Reinecke <[email protected]>

---
  drivers/scsi/virtio_scsi.c       | 62
++++++++++++++++++++++++++++++----------
  include/uapi/linux/virtio_scsi.h |  1 +

  2 files changed, 48 insertions(+), 15 deletions(-)

drivers/scsi/virtio_scsi.c +11 -5

index f925fbd..63c2c85 100644
@@ -356,8 +356,12 @@ static void
 virtscsi_handle_transport_reset(struct virtio_scsi *vscsi,
       struct scsi_device *sdev;
       struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
       unsigned int target = event->lun[1];
-    unsigned int lun = (event->lun[2] << 8) | event->lun[3];
+    u64 lun;
 
+    if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_NATIVE_LUN))
+        lun = scsilun_to_int((struct scsi_lun *)event->lun) >> 16;
+    else
+        lun = (event->lun[2] << 8) | event->lun[3];
 
 

@@ -524,10 +532,16 @@ static void virtio_scsi_init_hdr(struct
 virtio_device *vdev,
                    int target_id,
                    struct scsi_cmnd *sc)
   {
-    cmd->lun[0] = 1;
-    cmd->lun[1] = target_id;
-    cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
-    cmd->lun[3] = sc->device->lun & 0xff;
+    if (virtio_has_feature(vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+        u64 lun = sc->device->lun << 16;
+        lun |= ((u64)1 << 8) | (u64)target_id;
+        int_to_scsilun(lun, (struct scsi_lun *)&cmd->lun);
+    } else {
+        cmd->lun[0] = 1;

drivers/scsi/virtio_scsi.c static void virtio_scsi_init_hdr(struct virtio_device *vdev, · line 544

Quoted 4 lines

drivers/scsi/virtio_scsi.c +4 -0

+        cmd->lun[1] = target_id;
+        cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
+        cmd->lun[3] = sc->device->lun & 0xff;
+    }

Above 2 patterns seem to repeat. Have helper functions (similar to int_to_scsilun()) now that it's more than just 4 lines of filling in the virtio lun?

drivers/scsi/virtio_scsi.c static int virtscsi_abort(struct scsi_cmnd *sc) · line 874

Quoted 5 lines

drivers/scsi/virtio_scsi.c +0 -0

@@ -851,10 +871,18 @@ static int virtscsi_abort(struct scsi_cmnd *sc)
           .subtype = VIRTIO_SCSI_T_TMF_ABORT_TASK,
 
           .lun[0] = 1,
           .lun[1] = target_id,

drop those 2 superfluous lines, too?

Quoted 18 lines

drivers/scsi/virtio_scsi.c +10 -2

-        .lun[2] = (sc->device->lun >> 8) | 0x40,
-        .lun[3] = sc->device->lun & 0xff,
           .tag = cpu_to_virtio64(vscsi->vdev, (unsigned long)sc),
       };
+    if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+        u64 lun = sc->device->lun << 16;
+        lun |= ((u64)1 << 8) | (u64)target_id;
+        int_to_scsilun(lun, (struct scsi_lun *)&cmd->req.tmf.lun);
+    } else {
+        cmd->req.tmf.lun[0] = 1;
+        cmd->req.tmf.lun[1] = target_id;
+        cmd->req.tmf.lun[2] = (sc->device->lun >> 8) | 0x40;
+        cmd->req.tmf.lun[3] = sc->device->lun & 0xff;
+    }
 
       return virtscsi_tmf(vscsi, cmd);

  }

drivers/scsi/virtio_scsi.c static int virtscsi_probe(struct virtio_device · line 1460

Quoted 5 lines

drivers/scsi/virtio_scsi.c +0 -0

@@ -1429,7 +1457,10 @@ static int virtscsi_probe(struct virtio_device
 *vdev)
       /* LUNs > 256 are reported with format 1, so they go in the range
        * 16640-32767.
        */

Above old comment now only seems to apply to the then case of the following if statement, not to the else case.

Quoted 15 lines · Steffen Maier

Quoted 7 lines · Hannes Reinecke

The 'native LUN' feature allows virtio-scsi to pass in the LUN numbers from the underlying storage directly, without having to modify the LUN number itself.
It works by shifting the existing LUN number down by 8 bytes, and add the virtio-specific 8-byte LUN steering header. With that virtio doesn't have to mangle the LUN number, allowing us to pass the 'real' LUN number to the guest.

I only see shifts by 16 bits in the code below which would be 2 bytes. I had a quick look at the corresponding qemu code which looked the same to me.
What's the relation to 8 byte shifting, which would be 64 bit shift and thus odd for a 64 bit LUN, mentioned in the description here?

Quoted 3 lines

If the code keeps the LUN level 1 and 2 (dropping level 3 and 4) and I just don't understand it, it would be fine, I guess.

Yeah, messed that one up. It should be 8 _bits_, obviously.

Quoted 6 lines

Quoted 3 lines

Of course, we do cut off the last 8 bytes of the 'real' LUN number, but I'm not aware of any array utilizing that, so the impact should be negligible.

Why did we do v3.17 commit 9cb78c16f5da ("scsi: use 64-bit LUNs")? ;-)

Because that patch just lifts the internal code to use 64-bit LUNs without any changes to the behaviour.
This one uses the internal 64-bit LUNs and actually changes the behaviour.

Quoted 22 lines

Quoted 22 lines

Signed-off-by Hannes Reinecke <[email protected]>

---
   drivers/scsi/virtio_scsi.c       | 62
++++++++++++++++++++++++++++++----------
   include/uapi/linux/virtio_scsi.h |  1 +
   2 files changed, 48 insertions(+), 15 deletions(-)

drivers/scsi/virtio_scsi.c +6 -4

@@ -524,10 +532,16 @@ static void virtio_scsi_init_hdr(struct
 virtio_device *vdev,
                    int target_id,
                    struct scsi_cmnd *sc)
   {
-    cmd->lun[0] = 1;
-    cmd->lun[1] = target_id;
-    cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
-    cmd->lun[3] = sc->device->lun & 0xff;
+    if (virtio_has_feature(vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+        u64 lun = sc->device->lun << 16;
+        lun |= ((u64)1 << 8) | (u64)target_id;
+        int_to_scsilun(lun, (struct scsi_lun *)&cmd->lun);
+    } else {
+        cmd->lun[0] = 1;

Quoted 9 lines

drivers/scsi/virtio_scsi.c static void virtio_scsi_init_hdr(struct virtio_device *vdev, · line 544

Quoted 4 lines

drivers/scsi/virtio_scsi.c +4 -0

+        cmd->lun[1] = target_id;
+        cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
+        cmd->lun[3] = sc->device->lun & 0xff;
+    }

Above 2 patterns seem to repeat. Have helper functions (similar to int_to_scsilun()) now that it's more than just 4 lines of filling in the virtio lun?

Yes, can do.

Quoted 27 lines · Hannes Reinecke

Quoted 17 lines · Steffen Maier

Quoted 15 lines · Hannes Reinecke

drivers/scsi/virtio_scsi.c +6 -4

@@ -524,10 +532,16 @@ static void virtio_scsi_init_hdr(struct
 virtio_device *vdev,
                     int target_id,
                     struct scsi_cmnd *sc)
    {
-    cmd->lun[0] = 1;
-    cmd->lun[1] = target_id;
-    cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
-    cmd->lun[3] = sc->device->lun & 0xff;
+    if (virtio_has_feature(vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+        u64 lun = sc->device->lun << 16;
+        lun |= ((u64)1 << 8) | (u64)target_id;
+        int_to_scsilun(lun, (struct scsi_lun *)&cmd->lun);
+    } else {
+        cmd->lun[0] = 1;

Quoted 9 lines

drivers/scsi/virtio_scsi.c static void virtio_scsi_init_hdr(struct virtio_device *vdev, · line 544

Quoted 4 lines

drivers/scsi/virtio_scsi.c +4 -0

+        cmd->lun[1] = target_id;
+        cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
+        cmd->lun[3] = sc->device->lun & 0xff;
+    }

Above 2 patterns seem to repeat. Have helper functions (similar to int_to_scsilun()) now that it's more than just 4 lines of filling in the virtio lun?

Quoted 1 line

Yes, can do.

Meanwhile I think I realized why I had trouble understanding what the code does. I guess, I expected a conversion with int_to_scsilun() first, and then we would fill in the virtio-specific parts of magic-one and target-ID. You do it just the other way round, which is OK.

Say we have a 4 level 64-bit LUN and represent it in hex using placeholder hexdigits for the 4 levels like this:
0xL1L1L2L2L3L3L4L4
Its decimal SCSI LUN representation (in hex) is:
0xL4L4L3L3L2L2L1L1
Then you shift left by 16 bits (2 bytes, 1 LUN level), basically dropping the 4th level:
0xL3L3L2L2L1L10000
The steering header is 0x01TT where TT is the target ID.
You bitwise or the virtio-specific parts into the SCSI LUN representation:
0xL3L3L2L2L1L101TT
Finally you convert it into the 64-bit LUN representation:
0x01TTL1L1L2L2L3L3
  0123456789abcdef [char array indexes]

So we nicely have the virtio-specific parts at those array indexes where the virtio-scsi protocol expects them.
The usage of the other bytes is now of course different from the original LUN encoding: It allows more than just peripheral and flat space addressing for the 1st level; and it now also uses levels 2 and 3 which were previously always zero. The 3rd level really requires 64-bit support in the kvm host kernel. This also means that a 4-level LUN is not supported unless we would create a new virtio-scsi protocol version that would transfer the target ID in a separate field not as part of the LUN field.

Did I get that right?

A similar explanation in a kernel doc comment for the helper conversion function(s) might be helpful.

Quoted 16 lines · Hannes Reinecke

Quoted 15 lines · Steffen Maier

Quoted 7 lines · Hannes Reinecke

The 'native LUN' feature allows virtio-scsi to pass in the LUN numbers from the underlying storage directly, without having to modify the LUN number itself.
It works by shifting the existing LUN number down by 8 bytes, and add the virtio-specific 8-byte LUN steering header. With that virtio doesn't have to mangle the LUN number, allowing us to pass the 'real' LUN number to the guest.

I only see shifts by 16 bits in the code below which would be 2 bytes. I had a quick look at the corresponding qemu code which looked the same to me.
What's the relation to 8 byte shifting, which would be 64 bit shift and thus odd for a 64 bit LUN, mentioned in the description here?

Quoted 4 lines

Quoted 3 lines

If the code keeps the LUN level 1 and 2 (dropping level 3 and 4) and I just don't understand it, it would be fine, I guess.

Yeah, messed that one up. It should be 8 _bits_, obviously.

Isn't it 16 bits or 2 bytes corresponding to one LUN level? See also below.

Quoted 10 lines

Quoted 6 lines

Quoted 3 lines

Of course, we do cut off the last 8 bytes of the 'real' LUN number, but I'm not aware of any array utilizing that, so the impact should be negligible.

Why did we do v3.17 commit 9cb78c16f5da ("scsi: use 64-bit LUNs")? ;-)

Because that patch just lifts the internal code to use 64-bit LUNs without any changes to the behaviour.
This one uses the internal 64-bit LUNs and actually changes the behaviour.

Sure, I was just being ironic, because your description sounded a bit as if all the LUN range extension is not even required because no storage array uses it.

Quoted 31 lines

Quoted 22 lines

Quoted 22 lines

Signed-off-by Hannes Reinecke <[email protected]>

---
   drivers/scsi/virtio_scsi.c       | 62
++++++++++++++++++++++++++++++----------
   include/uapi/linux/virtio_scsi.h |  1 +

   2 files changed, 48 insertions(+), 15 deletions(-)

drivers/scsi/virtio_scsi.c +6 -4

@@ -524,10 +532,16 @@ static void virtio_scsi_init_hdr(struct
 virtio_device *vdev,
                     int target_id,
                     struct scsi_cmnd *sc)
    {
-    cmd->lun[0] = 1;
-    cmd->lun[1] = target_id;
-    cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
-    cmd->lun[3] = sc->device->lun & 0xff;
+    if (virtio_has_feature(vdev, VIRTIO_SCSI_F_NATIVE_LUN)) {
+        u64 lun = sc->device->lun << 16;
+        lun |= ((u64)1 << 8) | (u64)target_id;
+        int_to_scsilun(lun, (struct scsi_lun *)&cmd->lun);
+    } else {
+        cmd->lun[0] = 1;

Quoted 9 lines

drivers/scsi/virtio_scsi.c static void virtio_scsi_init_hdr(struct virtio_device *vdev, · line 544

Quoted 4 lines

drivers/scsi/virtio_scsi.c +4 -0

+        cmd->lun[1] = target_id;
+        cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
+        cmd->lun[3] = sc->device->lun & 0xff;
+    }

Above 2 patterns seem to repeat. Have helper functions (similar to int_to_scsilun()) now that it's more than just 4 lines of filling in the virtio lun?

Quoted 1 line

Yes, can do.

Meanwhile I think I realized why I had trouble understanding what the code does. I guess, I expected a conversion with int_to_scsilun() first, and then we would fill in the virtio-specific parts of magic-one and target-ID.
You do it just the other way round, which is OK.

Say we have a 4 level 64-bit LUN and represent it in hex using placeholder hexdigits for the 4 levels like this:
0xL1L1L2L2L3L3L4L4
Its decimal SCSI LUN representation (in hex) is:
0xL4L4L3L3L2L2L1L1
Then you shift left by 16 bits (2 bytes, 1 LUN level), basically dropping the 4th level:
0xL3L3L2L2L1L10000
The steering header is 0x01TT where TT is the target ID.
You bitwise or the virtio-specific parts into the SCSI LUN representation:
0xL3L3L2L2L1L101TT
Finally you convert it into the 64-bit LUN representation:
0x01TTL1L1L2L2L3L3
  0123456789abcdef [char array indexes]
So we nicely have the virtio-specific parts at those array indexes where the virtio-scsi protocol expects them.
The usage of the other bytes is now of course different from the original LUN encoding: It allows more than just peripheral and flat space addressing for the 1st level; and it now also uses levels 2 and 3 which were previously always zero. The 3rd level really requires 64-bit support in the kvm host kernel.
This also means that a 4-level LUN is not supported unless we would create a new virtio-scsi protocol version that would transfer the target ID in a separate field not as part of the LUN field.

Did I get that right?

A similar explanation in a kernel doc comment for the helper conversion function(s) might be helpful.