Messages in this thread | | | Date | Wed, 17 Aug 2022 17:05:27 +0200 | From | Christian Brauner <> | Subject | Re: [PATCH] binderfs: rework superblock destruction |
| |
On Wed, Aug 17, 2022 at 03:32:03PM +0100, Al Viro wrote: > On Wed, Aug 17, 2022 at 03:19:13PM +0100, Al Viro wrote: > > On Wed, Aug 17, 2022 at 04:01:49PM +0200, Christian Brauner wrote: > > > On Wed, Aug 17, 2022 at 02:59:02PM +0100, Al Viro wrote: > > > > On Wed, Aug 17, 2022 at 03:03:06PM +0200, Christian Brauner wrote: > > > > > > > > > +static void binderfs_kill_super(struct super_block *sb) > > > > > +{ > > > > > + struct binderfs_info *info = sb->s_fs_info; > > > > > + > > > > > + if (info && info->ipc_ns) > > > > > + put_ipc_ns(info->ipc_ns); > > > > > + > > > > > + kfree(info); > > > > > + kill_litter_super(sb); > > > > > +} > > > > > > > > Other way round, please - shut the superblock down, *then* > > > > free the objects it'd been using. IOW, > > > > > > I wondered about that but a lot of places do it the other way around. > > > So maybe the expected order should be documented somewhere. > > > > ??? > > > > "If you are holding internal references to dentries/inodes/etc., drop them > > first; if you are going to free something that is used by filesystem > > methods, don't do that before the filesystem is shut down" > > > > That's just common sense... Which filesystems are doing that "the other > > way around"? > > Note that something like e.g. ramfs, where we have a dynamically allocated > object ->s_fs_info is pointing to and gets freed early in their ->kill_sb() > is somewhat misleading - it's used only for two things, one is the > creation of root directory inode (obviously not going to happen at any > point after mount) and another - ->show_options(). By the point we get > around to killing a superblock, it would better *NOT* have mounts pointing > to it that might show up in /proc/mounts and make us call ->show_options(). > > So there we really know that nothing during the shutdown will even look > at that thing we'd just freed. Not that there'd ever been a point allocating > it - all that object contains is one unsigned short, so we might as well > just have stored (void *)root_mode in ->s_fs_info. Oh, well...
Binderfs was really the first fs I ever wrote and back then I was trying to be as close to best practice at possible. One thing I remember being unclear about was what the best practice for filesystem shutdown would be. That included ->put_super() vs just ->kill_sb() but also the order in which kill_litter_super() and sb->s_fs_info cleanup should happen.
For binderfs the order does matter and that's also the reason I originally decided to use ->put_super() as it's called after evict_inodes() and gives the required ordering.
| |