[scsi] Retry TEST UNIT READY command

The TEST UNIT READY command is issued automatically when the device is
opened, and is not the result of a command being issued by the caller.
This is required in order that a permanent TEST UNIT READY failure can
be used to identify unusable paths in a multipath SAN device.

Since the TEST UNIT READY command is not part of the caller's command
issuing process, it is not covered by any external retry loops (such
as the main retry loop in sandev_command()).

We must therefore be prepared to retry the TEST UNIT READY command
within the SCSI layer itself.  We retry only the TEST UNIT READY
command so as not to multiply the number of potential retries for
normal commands (which are already retried by sandev_command()).

Signed-off-by: Michael Brown <mcb30@ipxe.org>

src/drivers/block/scsi.c

@@ -40,6 +40,9 @@ FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
  *
  */
 
+/** Maximum number of TEST UNIT READY retries */
+#define SCSI_READY_MAX_RETRIES 10
+
 /* Error numbers generated by SCSI sense data */
 #define EIO_NO_SENSE __einfo_error ( EINFO_EIO_NO_SENSE )
 #define EINFO_EIO_NO_SENSE \
@@ -240,6 +243,8 @@ struct scsi_device {
 	struct interface ready;
 	/** TEST UNIT READY process */
 	struct process process;
+	/** TEST UNIT READY retry count */
+	unsigned int retries;
 
 	/** List of commands */
 	struct list_head cmds;
@@ -875,20 +880,40 @@ static struct interface_descriptor scsidev_block_desc =
 static void scsidev_ready ( struct scsi_device *scsidev, int rc ) {
 
 	/* Shut down interface */
-	intf_shutdown ( &scsidev->ready, rc );
+	intf_restart ( &scsidev->ready, rc );
 
-	/* Close device on failure */
-	if ( rc != 0 ) {
-		DBGC ( scsidev, "SCSI %p not ready: %s\n",
-		       scsidev, strerror ( rc ) );
-		scsidev_close ( scsidev, rc );
+	/* Mark device as ready, if applicable */
+	if ( rc == 0 ) {
+		DBGC ( scsidev, "SCSI %p unit is ready\n", scsidev );
+		scsidev->flags |= SCSIDEV_UNIT_READY;
+		xfer_window_changed ( &scsidev->block );
+		return;
+	}
+	DBGC ( scsidev, "SCSI %p not ready: %s\n", scsidev, strerror ( rc ) );
+
+	/* SCSI targets have an annoying habit of returning occasional
+	 * pointless "error" messages such as "power-on occurred", so
+	 * we have to be prepared to retry commands.
+	 *
+	 * For most commands, we rely on the caller (e.g. the generic
+	 * SAN device layer) to retry commands as needed.  However, a
+	 * TEST UNIT READY failure is used as an indication that the
+	 * whole SCSI device is unavailable and should be closed.  We
+	 * must therefore perform this retry loop within the SCSI
+	 * layer.
+	 */
+	if ( scsidev->retries++ < SCSI_READY_MAX_RETRIES ) {
+		DBGC ( scsidev, "SCSI %p retrying (retry %d)\n",
+		       scsidev, scsidev->retries );
+		scsidev->flags &= ~SCSIDEV_UNIT_TESTED;
+		process_add ( &scsidev->process );
 		return;
 	}
 
-	/* Mark device as ready */
-	scsidev->flags |= SCSIDEV_UNIT_READY;
-	xfer_window_changed ( &scsidev->block );
-	DBGC ( scsidev, "SCSI %p unit is ready\n", scsidev );
+	/* Close device */
+	DBGC ( scsidev, "SCSI %p never became ready: %s\n",
+	       scsidev, strerror ( rc ) );
+	scsidev_close ( scsidev, rc );
 }
 
 /** SCSI device TEST UNIT READY interface operations */