This is based on ideas from https://lkml.kernel.org/r/20220128130022.1750906-1-hch@xxxxxx but also kills "disk->open_mutex => lo->lo_mutex" dependency. Since __loop_clr_fd() runs synchronously from lo_release(), the reported regressions should be fixed. We can apply Christoph's [PATCH 1/8] loop: de-duplicate the idle worker freeing code [PATCH 2/8] loop: initialize the worker tracking fields once [PATCH 3/8] loop: remove the racy bd_inode->i_mapping->nrpages asserts patches after this series. I'm surprised with a lot of "if (!release)" usage in __loop_clr_fd() needed for avoid waiting for I/O request. By the way, does bdev_disk_changed() from __loop_clr_fd() involve I/O request which we are trying to avoid?
