RE: [PATCH v2 2/2] hfs: Update sanity check of the root record

[Viacheslav Dubeyko]

On Tue, 2025-11-11 at 00:23 +0000, George Anthony Vernon wrote:
> On Mon, Nov 10, 2025 at 11:34:39PM +0000, Viacheslav Dubeyko wrote:
> > On Mon, 2025-11-10 at 23:03 +0000, George Anthony Vernon wrote:
> > > On Tue, Nov 04, 2025 at 11:01:31PM +0000, Viacheslav Dubeyko wrote:
> > > > On Tue, 2025-11-04 at 01:47 +0000, George Anthony Vernon wrote:
> > > > > syzbot is reporting that BUG() in hfs_write_inode() fires upon unmount
> > > > > operation when the inode number of the record retrieved as a result of
> > > > > hfs_cat_find_brec(HFS_ROOT_CNID) is not HFS_ROOT_CNID, for commit
> > > > > b905bafdea21 ("hfs: Sanity check the root record") checked the record
> > > > > size and the record type but did not check the inode number.
> > > > > 
> > > > > Reported-by: syzbot+97e301b4b82ae803d21b@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > > > Closes: https://syzkaller.appspot.com/bug?extid=97e301b4b82ae803d21b      
> > > > > Signed-off-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
> > > > > Signed-off-by: George Anthony Vernon <contact@xxxxxxxxxxx>
> > > > > ---
> > > > >  fs/hfs/super.c | 2 +-
> > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/fs/hfs/super.c b/fs/hfs/super.c
> > > > > index 47f50fa555a4..a7dd20f2d743 100644
> > > > > --- a/fs/hfs/super.c
> > > > > +++ b/fs/hfs/super.c
> > > > > @@ -358,7 +358,7 @@ static int hfs_fill_super(struct super_block *sb, struct fs_context *fc)
> > > > >  			goto bail_hfs_find;
> > > > >  		}
> > > > >  		hfs_bnode_read(fd.bnode, &rec, fd.entryoffset, fd.entrylength);
> > > > > -		if (rec.type != HFS_CDR_DIR)
> > > > > +		if (rec.type != HFS_CDR_DIR || rec.dir.DirID != cpu_to_be32(HFS_ROOT_CNID))
> > > > 
> > > > This check is completely unnecessary. Because, we have hfs_iget() then [1]:
> > > > 
> > > > The hfs_iget() calls iget5_locked() [2]:
> > > > 
> > > > And iget5_locked() calls hfs_read_inode(). And hfs_read_inode() will call
> > > > is_valid_cnid() after applying your patch. So, is_valid_cnid() in
> > > > hfs_read_inode() can completely manage the issue. This is why we don't need in
> > > > this modification after your first patch.
> > > > 
> > > 
> > > I think Tetsuo's concern is that a directory catalog record with
> > > cnid > 15 might be returned as a result of hfs_bnode_read, which
> > > is_valid_cnid() would not protect against. I've satisfied myself that
> > > hfs_bnode_read() in hfs_fill_super() will populate hfs_find_data fd
> > > correctly and crash out if it failed to find a record with root CNID so
> > > this path is unreachable and there is no need for the second patch.
> > > 
> > 
> > Technically speaking, we can adopt this check to be completely sure that nothing
> > will be wrong during the mount operation. But I believe that is_valid_cnid()
> > should be good enough to manage this. Potential argument could be that the check
> > of rec.dir.DirID could be faster operation than to call hfs_iget(). But mount is
> > rare and not very fast operation, anyway. And if we fail to mount, then the
> > speed of mount operation is not very important.
> 
> Agreed we're not worried about speed that the mount operation can reach
> fail case. The check would have value if the bnode populated in
> hfs_find_data fd by hfs_cat_find_brec() is bad. That would be very
> defensive, I'm not sure it's necessary.
> 
> Maybe is_valid_cnid() should be is_valid_catalog_cnid(), since that is
> what it is actually testing for at the interface with the VFS. Would you
> agree?
> 

CNID is abbreviation of Catalog Node ID. So, is_valid_catalog_cnid() will sound
like Catalog Catalog Node ID. :)

Thanks,
Slava.

> > 
> > > > But I think we need to check that root_inode is not bad inode afterwards:
> > > > 
> > > > 	root_inode = hfs_iget(sb, &fd.search_key->cat, &rec);
> > > > 	hfs_find_exit(&fd);
> > > > 	if (!root_inode || is_bad_inode(root_inode))
> > > > 		goto bail_no_root;
> > > 
> > > Agreed, I see hfs_read_inode might return a bad inode. Thanks for
> > > catching this. I noticed also that it returns an int but the return
> > > value holds no meaning; it is always zero.
> > > 
> > > 
> > 
> > I've realized that hfs_write_inode() doesn't check that inode is bad like other
> > file systems do. Probably, we need to have this check too.
> 
> Good point, and similarly with HFS+. I'll take a look.
> 
> Thanks,
> 
> George