On 2022/03/29 18:42, Jan Kara wrote:

Ah, OK, so you'd always set LO_AUTOCLEAR and leave cleanup to happen
from lo_release()? That makes sense to me.

"loop: remove lo_refcount and avoid lo_mutex in ->open / ->release" is going to
make "not always" set LO_AUTOCLEAR due to lack of disk->open_mutex serialization.

That's why this topic is discussed.

We could use loop_global_lock_killable()/loop_global_unlock() in order to compensate
for lack of disk->open_mutex serialization, as described in the bottom half of
https://lkml.kernel.org/r/03628e13-ca56-4ed0-da5a-ee698c83f48d@I-love.SAKURA.ne.jp .