On Tue, Dec 05, 2023 at 11:16:33PM -0800, Christoph Hellwig wrote: > On Tue, Dec 05, 2023 at 10:44:30PM -0800, Eric Biggers wrote: > > There are lots of filesystems that free their ->s_fs_info in ->put_super(). Are > > they all wrong? > > Wrong is the wrong word :) > > For simple file systems that either don't use block devices, or just > the main one set up by the super.c code using ->put_super is perfectly > fine. Once a file system does something complicated like setting up > their own block devices, or trying to reuse super blocks using custom > rules it basically has to free ->s_fs_info in it's own ->kill_sb > handler. This whole area is a bit messy unfortunately. > btrfs releases its block devices in ->put_super(), so with your proposal that would need to be moved to ->kill_sb() as well, right? I guess I'm a bit concerned about introducing a requirement "->put_super() must not release the block devices" which seemingly didn't exist before. - Eric
