[RFC] lsm: fs: Use i_callback to free i_security in RCU callback

[Song Liu]

On Tue, Dec 17, 2024 at 8:44 AM Paul Moore <paul at paul-moore.com> wrote:
>
> On Mon, Dec 16, 2024 at 8:24 PM Song Liu <song at kernel.org> wrote:
> > On Mon, Dec 16, 2024 at 4:22 PM Paul Moore <paul at paul-moore.com> wrote:
> > >
> > > On Mon, Dec 16, 2024 at 6:43 PM Song Liu <song at kernel.org> wrote:
> > > >
> > > > inode->i_security needes to be freed from RCU callback. A rcu_head was
> > > > added to i_security to call the RCU callback. However, since struct inode
> > > > already has i_rcu, the extra rcu_head is wasteful. Specifically, when any
> > > > LSM uses i_security, a rcu_head (two pointers) is allocated for each
> > > > inode.
> > > >
> > > > Add security_inode_free_rcu() to i_callback to free i_security so that
> > > > a rcu_head is saved for each inode. Special care are needed for file
> > > > systems that provide a destroy_inode() callback, but not a free_inode()
> > > > callback. Specifically, the following logic are added to handle such
> > > > cases:
> > > >
> > > >  - XFS recycles inode after destroy_inode. The inodes are freed from
> > > >    recycle logic. Let xfs_inode_free_callback() and xfs_inode_alloc()
> > > >    call security_inode_free_rcu() before freeing the inode.
> > > >  - Let pipe free inode from a RCU callback.
> > > >  - Let btrfs-test free inode from a RCU callback.
> > >
> > > If I recall correctly, historically the vfs devs have pushed back on
> > > filesystem specific changes such as this, requiring LSM hooks to
> > > operate at the VFS layer unless there was absolutely no other choice.
> > >
> > > From a LSM perspective I'm also a little concerned that this approach
> > > is too reliant on individual filesystems doing the right thing with
> > > respect to LSM hooks which I worry will result in some ugly bugs in
> > > the future.
> >
> > Totally agree with the concerns. However, given the savings is quite
> > significant (saving two pointers per inode), I think the it may justify
> > the extra effort to maintain the logic. Note that, some LSMs are
> > enabled in most systems and cannot be easily disabled, so I am
> > assuming most systems will see the savings.
>
> I suggest trying to find a solution that is not as fragile in the face
> of cross subsystem changes and ideally also limits the number of times
> the LSM calls must be made in individual filesystems.

There are three (groups of) subsystems here: VFS, file systems, and
LSM. It is not really possible to do this without crossing subsystem
boundaries. Specifically, since VFS allow a file system to have
destroy_inode callback, but not free_inode callback, we will need
such file systems to handle rcu callback. Does this make sense?

Suggestions on how we can solve this better are always appreciated.

Thanks,
Song