Re: [PATCH] hugetlbfs: simplify destroying inode

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 4 Apr 2011 11:36:56 +0800
Hillf Danton <dhillf@gmail.com> wrote:

> Just before reclaimed by the hugetlbfs inode cache, little difference
> could be made by initializing the list head of the dentry member of
> vfs inode, thus the operation of initialization could be removed to
> simplify the destroy of hugetlbfs inode.
> 
> Signed-off-by: Hillf Danton <gmail.com>
> ---
> 
> --- a/fs/hugetlbfs/inode.c	2011-03-30 03:09:48.000000000 +0800
> +++ b/fs/hugetlbfs/inode.c	2011-04-04 11:27:30.000000000 +0800
> @@ -665,7 +665,6 @@ static struct inode *hugetlbfs_alloc_ino
>  static void hugetlbfs_i_callback(struct rcu_head *head)
>  {
>  	struct inode *inode = container_of(head, struct inode, i_rcu);
> -	INIT_LIST_HEAD(&inode->i_dentry);
>  	kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
>  }

How well was this tested?

hugetlbfs_inode_cachep has a slab constructor,
init_once()->inode_init_once().  The slab protocol requires that
objects be returned to the cache in the same state as that which the
constructor creates.  (It's a bit weird and it would probably be faster
and certainly saner if the ctor were run by kmem_cache_alloc()).

inode_init_once() does INIT_LIST_HEAD(&inode->i_dentry), so objects
should be returned to the cache in that state.  If no other code was
already reliably reinitialising ->i_dentry, the patch will add a bug.