On Wed, Sep 01, 2021 at 01:19:44PM +0900, Tetsuo Handa wrote:

Are the changes by add_disk() made atomically against the caller?
If there is a moment where "struct block_device_operations"->open can be
called when add_disk() might still fail, how is it protected from the
kfree() path?

The new add_disk is structured so tht it won't fail after the block
device inode is hashed.  That is, the last possible failure is before
->open can be called.