Re: [PATCH v3] hfs: give the internal B-tree inodes a type and keep catalog records off them

[Viacheslav Dubeyko <slava@dubeyko.com>]

On Tue, 2026-10-13 at 14:44 +0800, yangguikun wrote:
> hfs_btree_open() builds the extents and catalog B-tree inodes with a
> raw
> iget_locked() and never assigns them a file type, so both stay cached
> with
> i_mode == 0 after every mount.  Two problems follow from that, and
> both are
> reachable from a crafted image.
> 
> First, a catalog record that names CNID 3 or 4 makes hfs_lookup() ->
> hfs_iget()
> return the cached B-tree inode - hfs_test_inode() matches on i_ino
> alone - and
> the typeless inode then reaches the generic open path, where it trips
> VFS_BUG_ON_INODE() in may_open() with CONFIG_DEBUG_VFS=y:
> 
>   VFS_BUG_ON_INODE(!IS_ANON_FILE(inode)): inode:... fs:hfs mode:0
> opflags:0xc
>   ------------[ cut here ]------------
>   kernel BUG at fs/namei.c:4271!
>   RIP: 0010:may_open+0x2ba/0x480 fs/namei.c:4271
>   Call Trace:
>    do_open fs/namei.c:4698 [inline]
>    path_openat+0x1600/0x3de0 fs/namei.c:4863
>    __x64_sys_openat+0x13f/0x1f0 fs/open.c:1385
> 
> Without CONFIG_DEBUG_VFS the assertion compiles away and the same
> image
> silently yields an openable typeless inode, so this is not merely a
> debug-config nuisance.
> 
> Second, once the inode has a type, such a record still hands the
> internal
> B-tree file out as a regular dentry: it can be opened and read, which
> exposes
> the B-tree, and from then on the VFS treats an inode that has no
> catalog
> record behind it as an ordinary file.  hfs_is_valid_cnid() does not
> close this
> path: it runs from hfs_read_inode(), and a lookup that hits the
> already cached
> B-tree inode never reaches that function.
> 
> Fix both.  Add hfs_system_read_inode(), which reads the fork from the
> MDB and
> assigns S_IFREG - the counterpart of hfsplus_system_read_inode() -
> and
> hfs_system_iget(), which looks the inode up through the normal inode-
> cache path
> and can fail before the inode is published.  hfs_btree_open() now
> uses
> hfs_system_iget(), and the extent-record validation moves into the
> new read
> helper.  hfs_iget() refuses catalog records that name a reserved CNID
> below
> HFS_FIRSTUSER_CNID; the root directory, the one reserved CNID that
> legitimately
> comes through there, is exempt.
> 
> Found by syzkaller-style fuzzing of v7.2-rc7; still reproducible on
> current
> mainline.
> 
> Suggested-by: Viacheslav Dubeyko <slava@dubeyko.com>
> Link:
> https://lore.kernel.org/linux-fsdevel/tencent_24D928451FEFA8A35482EFFA9132A7305307@qq.com/
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Cc: stable@vger.kernel.org
> Signed-off-by: yangguikun <yangguikun19@qq.com>
> ---
>  fs/hfs/btree.c  | 39 +++------------------------
>  fs/hfs/hfs_fs.h |  1 +
>  fs/hfs/inode.c  | 82
> +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  3 files changed, 86 insertions(+), 36 deletions(-)
> 
> diff --git a/fs/hfs/btree.c b/fs/hfs/btree.c
> index 41b4e8f..b5892de 100644
> --- a/fs/hfs/btree.c
> +++ b/fs/hfs/btree.c
> @@ -155,44 +155,11 @@ struct hfs_btree *hfs_btree_open(struct
> super_block *sb, u32 id, btree_keycmp ke
>  	tree->cnid = id;
>  	tree->keycmp = keycmp;
>  
> -	tree->inode = iget_locked(sb, id);
> -	if (!tree->inode)
> +	tree->inode = hfs_system_iget(sb, id);
> +	if (IS_ERR(tree->inode))
>  		goto free_tree;

I worried that we jump to free_tree now. Before, we cannot get inode
but now method can or cannot get inode. We need somehow distinguish
these cases now and to jump to free_tree or to free_inode. Or we need
to manage it in more smart way.

> -	BUG_ON(!(inode_state_read_once(tree->inode) & I_NEW));

I am not sure that it's good idea to drop this. If hfs_system_iget()
quietly returns an existing inode. In hfs_btree_open() that would mean
the same tree was opened twice.

> -	{
> -	struct hfs_mdb *mdb = HFS_SB(sb)->mdb;
> -	HFS_I(tree->inode)->flags = 0;
> -	mutex_init(&HFS_I(tree->inode)->extents_lock);
> -	switch (id) {
> -	case HFS_EXT_CNID:
> -		hfs_inode_read_fork(tree->inode, mdb->drXTExtRec,
> mdb->drXTFlSize,
> -				    mdb->drXTFlSize,
> be32_to_cpu(mdb->drXTClpSiz));
> -		if (HFS_I(tree->inode)->alloc_blocks >
> -					HFS_I(tree->inode)-
> >first_blocks) {
> -			pr_err("invalid btree extent records\n");
> -			unlock_new_inode(tree->inode);
> -			goto free_inode;
> -		}
>  
> -		tree->inode->i_mapping->a_ops = &hfs_btree_aops;
> -		break;
> -	case HFS_CAT_CNID:
> -		hfs_inode_read_fork(tree->inode, mdb->drCTExtRec,
> mdb->drCTFlSize,
> -				    mdb->drCTFlSize,
> be32_to_cpu(mdb->drCTClpSiz));
> -
> -		if (!HFS_I(tree->inode)->first_blocks) {
> -			pr_err("invalid btree extent records (0
> size)\n");
> -			unlock_new_inode(tree->inode);
> -			goto free_inode;
> -		}
> -
> -		tree->inode->i_mapping->a_ops = &hfs_btree_aops;
> -		break;
> -	default:
> -		BUG();
> -	}
> -	}
> -	unlock_new_inode(tree->inode);
> +	tree->inode->i_mapping->a_ops = &hfs_btree_aops;

 Set a_ops inside hfs_system_read_inode(). Right now tree->inode-
>i_mapping->a_ops = &hfs_btree_aops; runs in hfs_btree_open() after
unlock_new_inode(), so the inode is visible in the hash briefly without
its aops.

>  
>  	mapping = tree->inode->i_mapping;
>  	folio = filemap_grab_folio(mapping, 0);
> diff --git a/fs/hfs/hfs_fs.h b/fs/hfs/hfs_fs.h
> index e250f87..26505f7 100644
> --- a/fs/hfs/hfs_fs.h
> +++ b/fs/hfs/hfs_fs.h
> @@ -219,6 +219,7 @@ extern void hfs_inode_read_fork(struct inode
> *inode, struct hfs_extent *ext,
>  				u32 clump_size);
>  extern struct inode *hfs_iget(struct super_block *sb, struct
> hfs_cat_key *key,
>  				hfs_cat_rec *rec);
> +extern struct inode *hfs_system_iget(struct super_block *sb, u32
> cnid);
>  extern void hfs_evict_inode(struct inode *inode);
>  extern void hfs_delete_inode(struct inode *inode);
>  
> diff --git a/fs/hfs/inode.c b/fs/hfs/inode.c
> index 2aef3c3..5c00383 100644
> --- a/fs/hfs/inode.c
> +++ b/fs/hfs/inode.c
> @@ -421,6 +421,87 @@ static int hfs_read_inode(struct inode *inode,
> void *data)
>  	return 0;
>  }
>  
> +/**
> + * hfs_system_read_inode - read an internal B-tree inode
> + * @inode: inode to initialize, hashed by its CNID
> + *
> + * The extent and catalog B-trees are described by the MDB instead
> of by
> + * catalog records, so they do not go through hfs_read_inode().
> + *
> + * Return: 0 on success, -EIO if the MDB does not describe this
> CNID.
> + */
> +static int hfs_system_read_inode(struct inode *inode)
> +{
> +	struct hfs_sb_info *hsb = HFS_SB(inode->i_sb);
> +	struct hfs_mdb *mdb = hsb->mdb;
> +
> +	HFS_I(inode)->flags = 0;
> +	HFS_I(inode)->rsrc_inode = NULL;
> +	mutex_init(&HFS_I(inode)->extents_lock);
> +	HFS_I(inode)->tz_secondswest = sys_tz.tz_minuteswest * 60;
> +
> +	inode->i_uid = hsb->s_uid;
> +	inode->i_gid = hsb->s_gid;
> +	set_nlink(inode, 1);
> +
> +	switch (inode->i_ino) {
> +	case HFS_EXT_CNID:
> +		hfs_inode_read_fork(inode, mdb->drXTExtRec, mdb-
> >drXTFlSize,
> +				    mdb->drXTFlSize,
> be32_to_cpu(mdb->drXTClpSiz));
> +		if (HFS_I(inode)->alloc_blocks > HFS_I(inode)-
> >first_blocks) {
> +			pr_err("invalid btree extent records\n");
> +			return -EIO;
> +		}
> +		break;
> +	case HFS_CAT_CNID:
> +		hfs_inode_read_fork(inode, mdb->drCTExtRec, mdb-
> >drCTFlSize,
> +				    mdb->drCTFlSize,
> be32_to_cpu(mdb->drCTClpSiz));
> +		if (!HFS_I(inode)->first_blocks) {
> +			pr_err("invalid btree extent records (0
> size)\n");
> +			return -EIO;
> +		}
> +		break;
> +	default:
> +		return -EIO;
> +	}
> +
> +	/*
> +	 * may_open() requires a file type, and a crafted catalog
> record can
> +	 * reach this inode by its CNID.
> +	 */
> +	inode->i_mode = S_IFREG;
> +
> +	return 0;
> +}
> +
> +/**
> + * hfs_system_iget - look up an internal B-tree inode by CNID
> + * @sb: super block
> + * @cnid: HFS_EXT_CNID or HFS_CAT_CNID
> + *
> + * Return: the inode, or an ERR_PTR() if it cannot be initialized.
> + */
> +struct inode *hfs_system_iget(struct super_block *sb, u32 cnid)
> +{
> +	struct inode *inode;
> +	int err;
> +
> +	inode = iget_locked(sb, cnid);
> +	if (!inode)
> +		return ERR_PTR(-ENOMEM);
> +	if (!(inode_state_read_once(inode) & I_NEW))
> +		return inode;
> +
> +	err = hfs_system_read_inode(inode);
> +	if (err) {
> +		iget_failed(inode);
> +		return ERR_PTR(err);
> +	}
> +
> +	unlock_new_inode(inode);
> +	return inode;
> +}
> +
>  /*
>   * __hfs_iget()
>   *
> @@ -446,6 +527,15 @@ struct inode *hfs_iget(struct super_block *sb,
> struct hfs_cat_key *key, hfs_cat_
>  	default:
>  		return NULL;
>  	}
> +
> +	/*
> +	 * hfs_test_inode() matches on i_ino only, so a catalog
> record naming
> +	 * a reserved CNID would hand out a B-tree inode as a
> regular dentry.
> +	 * The root directory is the one reserved CNID that comes
> through here.
> +	 */
> +	if (cnid != HFS_ROOT_CNID && cnid < HFS_FIRSTUSER_CNID)

I am guessing maybe we need to introduce a special method:

static inline bool hfs_is_system_reserved_cnid(u32 cnid)
{
	if (likely(cnid > HFS_ROOT_CNID	&& cnid <
HFS_FIRSTUSER_CNID))
		return true;
	else
        	return false;
}

What do you think?

Thanks,
Slava.

> +		return NULL;
> +
>  	inode = iget5_locked(sb, cnid, hfs_test_inode,
> hfs_read_inode, &data);
>  	if (inode && (inode_state_read_once(inode) & I_NEW))
>  		unlock_new_inode(inode);