Hi,

On Tue, 2012-03-13 at 13:55 -0400, Bob Peterson wrote:
> Hi,
>
> This patch initializes the hash table cache when inodes are
> created in order to prevent using a hash table pointer left over
> from a previous inode. It also adds boundary checking on the
> hash table.
>
> Regards,
>
> Bob Peterson
> Red Hat File Systems
>
> Signed-off-by: Bob Peterson <rpeterso@redhat.com>
> --
> diff --git a/fs/gfs2/dir.c b/fs/gfs2/dir.c
> index c35573a..c2eb20f 100644
> --- a/fs/gfs2/dir.c
> +++ b/fs/gfs2/dir.c
> @@ -740,6 +740,7 @@ static int get_leaf_nr(struct gfs2_inode *dip, u32 index,
> hash = gfs2_dir_get_hash_table(dip);
> if (IS_ERR(hash))
> return PTR_ERR(hash);
> + BUG_ON(index >= (1 << dip->i_depth));
> *leaf_out = be64_to_cpu(*(hash + index));
> return 0;
> }
> diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c
> index 6172fa7..bd2abe3 100644
> --- a/fs/gfs2/super.c
> +++ b/fs/gfs2/super.c
> @@ -1576,6 +1576,7 @@ static struct inode *gfs2_alloc_inode(struct super_block *sb)
> ip->i_flags = 0;
> ip->i_gl = NULL;
> ip->i_rgd = NULL;
> + ip->i_hash_cache = NULL;
> }
> return &ip->i_inode;
> }
>

The question is how does this pointer land up being not NULL to start
with... I thought that you were looking into that from your earlier
comment in the bug,

Steve.

Hi,

On Wed, 2012-03-14 at 08:41 -0400, Bob Peterson wrote:
> ----- Original Message -----
> | Hi,
> |
> | On Tue, 2012-03-13 at 13:55 -0400, Bob Peterson wrote:
> | > Hi,
> | >
> | > This patch initializes the hash table cache when inodes are
> | > created in order to prevent using a hash table pointer left over
> | > from a previous inode. It also adds boundary checking on the
> | > hash table.
> | >
> | > Regards,
> | >
> | > Bob Peterson
> | > Red Hat File Systems
> | >
> | > Signed-off-by: Bob Peterson <rpeterso@redhat.com>
> | > --
> | > diff --git a/fs/gfs2/dir.c b/fs/gfs2/dir.c
> | > index c35573a..c2eb20f 100644
> | > --- a/fs/gfs2/dir.c
> | > +++ b/fs/gfs2/dir.c
> | > @@ -740,6 +740,7 @@ static int get_leaf_nr(struct gfs2_inode *dip,
> | > u32 index,
> | > hash = gfs2_dir_get_hash_table(dip);
> | > if (IS_ERR(hash))
> | > return PTR_ERR(hash);
> | > + BUG_ON(index >= (1 << dip->i_depth));
> | > *leaf_out = be64_to_cpu(*(hash + index));
> | > return 0;
> | > }
> | > diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c
> | > index 6172fa7..bd2abe3 100644
> | > --- a/fs/gfs2/super.c
> | > +++ b/fs/gfs2/super.c
> | > @@ -1576,6 +1576,7 @@ static struct inode *gfs2_alloc_inode(struct
> | > super_block *sb)
> | > ip->i_flags = 0;
> | > ip->i_gl = NULL;
> | > ip->i_rgd = NULL;
> | > + ip->i_hash_cache = NULL;
> | > }
> | > return &ip->i_inode;
> | > }
> | >
> |
> | The question is how does this pointer land up being not NULL to start
> | with... I thought that you were looking into that from your earlier
> | comment in the bug,
> |
> | Steve.
> Hi,
>
> My belief is that the i_hash_cache pointer was non-NULL because of
> a code path in the RHEL6 kernel that does not exist in the upstream
> kernel. (RHEL6 function gfs2_clear_inode was not invalidating the hash
> table cache like it should, but it doesn't exist in upstream GFS2).
>
> If my suspicion is correct, setting the i_hash_cache pointer to NULL
> from gfs2_alloc_inode is unnecessary and purely precautionary.
> I'm running some tests on the RHEL6 kernel to help prove that theory,
> but the test will probably take all day (it's a long-running test).
>
It would be better to add a test, if you are going to add something
here. We don't want to hide any future problem and there is no valid
reason for the pointer to be non-NULL at this point,

Steve.

> I can't reproduce the original problem with the upstream kernel
> because I can't run that test on my upstream box (it requires a
> special RHEL6 environment set up by a third party).
>
> Regards,
>
> Bob Peterson
> Red Hat File Systems