On Fri, Jan 10, 2014 at 07:53:41AM +0800, Linus Torvalds wrote:

On Fri, Jan 10, 2014 at 7:37 AM, Eric Paris [off-list ref] wrote:

quoted

but at least from an SELinux PoV, I think it's quick and easy, but wrong
for maintainability...

Yeah, it's a hack, and it's wrong, and we should figure out how to do
it right. Likely we should just tie the lifetime of the i_security
member directly to the lifetime of the inode itself, and just make the
rule be that security_inode_free() gets called from whatever frees the
inode itself, and *not* have an extra rcu callback etc. But that
sounds like a bigger change than I'm comfy with right now, so the
hacky one might be the band-aid to do for stable..

The problem, of course, is that all the different filesystems have
their own inode allocations/freeing. Of course, they all tend to share
the same pattern ("call_rcu xyz_i_callback"), so maybe we could try to
make that a more generic thing? Like have a "free_inode" vfs callback,
and do the call_rcu delaying at the VFS level..

And maybe, just maybe, we could just say that that is what
"destroy_inode()" is, and that we will just call it from rcu context.
All the IO has hopefully been done earlier  Yes/no?

Check what XFS is doing ;-/  That's where those call_rcu() have come from.
Sure, we can separate the simple "just do call_rcu(...->free_inode)" case
and hit it whenever full ->free_inode is there and ->destroy_inode isn't.
Not too pretty, but removal of tons of boilerplate might be worth doing
that anyway.  But ->destroy_inode() is still needed for cases where fs
has its own idea of inode lifetime rules.  Again, check what XFS is doing
in that area...

There's an extra source of headache, BTW - what about the "LSM stacking"
crowd and their plans?