When issuing IO handle broken, missing devices and failures

Previously we would issue the IO and assume everything
is going according to plan, this meant in the case of
failure to issue IO, we would hang forever. For example
if the device is removed.

Now we check the return code of IoCallDriver(), and handle
both direct failures, or sync calls (instead of expected
async call).

This gives us a better chance of detecting ejected, removed
or failed devices.

Signed-off-by: Jorgen Lundman <lundman@lundman.net>

Author: lundman

include/os/windows/zfs/sys/zfs_context_os.h

@@ -34,6 +34,7 @@
 		IRP *irp; \
 		void *b_addr; \
 		IO_STATUS_BLOCK IoStatus; \
+		boolean_t completion_called; \
 		PIO_WORKITEM work_item; \
 	} windows;
 

module/os/windows/zfs/vdev_disk.c

@@ -753,6 +753,8 @@ vdev_disk_io_intr(PDEVICE_OBJECT DeviceObject, PIRP irp, PVOID Context)
  */
 
 	VERIFY3P(zio->windows.work_item, !=, NULL);
+	zio->windows.completion_called = B_TRUE;
+
 	IoQueueWorkItem(zio->windows.work_item,
 	    (PIO_WORKITEM_ROUTINE)vdev_disk_io_start_done,
 	    DelayedWorkQueue, zio);
@@ -924,6 +926,7 @@ vdev_disk_io_start(zio_t *zio)
 	}
 
 	zio->windows.irp = irp;
+	zio->windows.completion_called = B_FALSE;
 
 	irpStack = IoGetNextIrpStackLocation(irp);
 
@@ -937,8 +940,18 @@ vdev_disk_io_start(zio_t *zio)
 	    TRUE,  // On Error
 	    TRUE); // On Cancel
 
-	IoCallDriver(dvd->vd_DeviceObject, irp);
+	NTSTATUS Status;
+	Status = IoCallDriver(dvd->vd_DeviceObject, irp);
+
+	// IO is in progress asynchronously...
+	if (Status == STATUS_PENDING)
+		return;
 
+	// IO completed synchronously or failed to start
+	// we clean up manually, assuming the completion
+	// callback was not called.
+	if (!zio->windows.completion_called)
+		vdev_disk_io_start_done(NULL, zio);
 }
 
 static void