Ocfs2 devel - Apr 2015 - vfs: Add general support to enforce project quota limits

Hello Dan,

On Thu 09-04-15 22:39:54, Dan Carpenter wrote:
> The patch 847aac644e92: "vfs: Add general support to enforce project
> quota limits" from Mar 19, 2015, leads to the following static
> checker warning:
  Thanks for letting us know.
> 	fs/ocfs2/quota_local.c:183 ocfs2_local_check_quota_file()
> 	error: buffer overflow 'lmagics' 2 <= 2
  OK, so the checker thinks that 'type' argument can be larger than
OCFS2_MAXQUOTAS.
> fs/ocfs2/quota_local.c
>    159  /* Check whether we understand format of quota files */
>    160  static int ocfs2_local_check_quota_file(struct super_block *sb, int
type)
>    161  {
>    162          unsigned int lmagics[OCFS2_MAXQUOTAS] =
OCFS2_LOCAL_QMAGICS;
>                                      ^^^^^^^^^^^^^^^
> This is 2.  Maybe the fix is to change this to MAXQUOTAS.
  No, that isn't the right fix. OCFS2 still supports only two quota types.
The right fix is to make sure that type == 2 cannot reach filesystems which
don't support it. See below...
>    163          unsigned int lversions[OCFS2_MAXQUOTAS] =
OCFS2_LOCAL_QVERSIONS;
>    164          unsigned int gmagics[OCFS2_MAXQUOTAS] =
OCFS2_GLOBAL_QMAGICS;
>    165          unsigned int gversions[OCFS2_MAXQUOTAS] =
OCFS2_GLOBAL_QVERSIONS;
>    166          unsigned int ino[OCFS2_MAXQUOTAS] = {
USER_QUOTA_SYSTEM_INODE,
>    167                                               
GROUP_QUOTA_SYSTEM_INODE };
>    168          struct buffer_head *bh = NULL;
>    169          struct inode *linode = sb_dqopt(sb)->files[type];
>    170          struct inode *ginode = NULL;
>    171          struct ocfs2_disk_dqheader *dqhead;
>    172          int status, ret = 0;
>    173  
>    174          /* First check whether we understand local quota file */
>    175          status = ocfs2_read_quota_block(linode, 0, &bh);
>    176          if (status) {
>    177                  mlog_errno(status);
>    178                  mlog(ML_ERROR, "failed to read quota file
header (type=%d)\n",
>    179                          type);
>    180                  goto out_err;
>    181          }
>    182          dqhead = (struct ocfs2_disk_dqheader *)(bh->b_data);
>    183          if (le32_to_cpu(dqhead->dqh_magic) != lmagics[type]) {
>                                                       ^^^^^^^^^^^^^
> This is one past the end of the array.  It used to be limitied in
> do_quotactl().
> 
> 	if (type >= (XQM_COMMAND(cmd) ? XQM_MAXQUOTAS : MAXQUOTAS))
> 
> The old logic was the XFS had 3 quotas and everyone else had 2 but now
> we raised MAXQUOTAS to 3 as well.
  Yes, but we have there a test:
        if (!(sb->s_quota_types & (1 << type)))
                return -EINVAL;
  which also limits allowed types. And OCFS2 sets this to QTYPE_MASK_USR |
QTYPE_MASK_GRP so type == 2 cannot reach it via quotactl(). So either this
is a false positive or there's some other path how type == 2 can reach
OCFS2 which I'm missing and I'd definitely like to learn about it. Can
you
investigate a bit please? Thanks.

								Honza
-- 
Jan Kara <jack at suse.cz>
SUSE Labs, CR

Ah.  Thanks for looking into it.  Smatch can't understand bit flags like
that yet.  And it also still doesn't track pointers very well.
Eventually those things will be improved.

I see now that this out of bounds is also prevented in
vfs_load_quota_inode().  We don't allow "(type == PRJQUOTA &&
sb->dq_op->get_projid == NULL)".  Since, ->get_projid() is always
NULL
for OCFS2 that means "type" can't equal to PRJQUOTA (2).

Hm..  is ->get_projid set anywhere?  Cscope and smatch can't find it if
it is.

I haven't investigated v1_read_file_info() but presumably it's the same.

regards,
dan carpenter