Re: [PATCH] vfs: move ACL cache lookup into generic code

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, Jul 23, 2011 at 04:29:44AM +0100, Al Viro wrote:

> OK, unless there are serious objections, I
> 	* apply Linus' patch as-is
> 	* fix that ocfs2 leak
> 	* replace posix_acl_..._masq() with saner helpers (take original
> acl + other arguments, return modified clone or ERR_PTR) and kill open-coded
> instances...

Done (along with 9p leak fixes).  NOTE: this is completely untested.
posix_acl_clone()/posix_acl_create_masq()/posix_acl_chmod_masq() all
became static in fs/posix_acl.c; replacement functions are
posix_acl_create() and posix_acl_chmod().  All callers converted, with
about a hundred lines of boilerplates gone.

And yes, it's pure copy-on-write.  The objects are refcounted, shared a lot
and the only primitives that did modifications had been *too* primitive -
their callers all started with creating a copy, then modifying it, then
letting somebody see the result.

Conclusion: that's an obvious free-by-RCU fodder.  I have not done that yet,
but it should be trivial to add.  posix_acl_release() would, instead of
kfree(), do kfree_rcu() (and we'd need to add rcu_head to struct posix_acl).

I'm not sure if merging the copying into ..._masq() would be worth doing.
posix_acl_clone() is basically kmemdup() and ..._masq() is a loop over all
array elements, with switch by element tag, modifying some of them.  Might
be worth merging with copying the elements...  Hell knows.

Anyway, that'll have to wait for tomorrow; I'm going down right now.  This
stuff (plus Tim's "mount lock scalability for internal mounts" patch) is
in #untested in usual place.  Comments/testing/etc. are welcome...