On 2021/12/21 4:16, Luis Chamberlain wrote:
> The kobject_alive() tells us if at least the device_add() had the
> kobject_add() complete.
Testing with error injection
@@ -3284,6 +3284,9 @@ int device_add(struct device *dev)
if (!dev)
goto done;
+ if (!strcmp(current->comm, "a.out"))
+ goto done;
+
if (!dev->p) {
error = device_private_init(dev);
if (error)
told me that kref count is 1 when reaching the out_disk_release_events label.
Thus,
if (!kobject_alive(&ddev->kobj))
seems wrong.
Christoph proposed deferring disk_alloc_events(). If it is safe to defer
disk_alloc_events(), that can be a fix. But I don't know if there is a path
which can invoke disk event functions (e.g. disk_block_events() from
blkdev_get_by_dev()) between device_add() and disk_alloc_events().
I didn't find a path, but at least the device file and some of sysfs files
are already visible before disk_alloc_events() is called...
