Hi,
     It appears that uninitialized r[3] in 
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for 
a particular case.

This sequence is below:

        xfs_bmap_add_extent_delay_real (
        ...
        xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
          /*
           * Filling in the middle part of a previous delayed allocation.
           * Contiguity is impossible here.
           * This case is avoided almost all the time.
           */
temp = new->br_startoff - PREV.br_startoff;
xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff;
r[1].br_blockcount = temp2;
xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK);
ip->i_df.if_lastex = idx + 1;
ip->i_d.di_nextents++;

Look at extent r[1]. It does not set br_startblock. That is, it is any 
random value. Now, look at the xfs_bmbt_set_all. Though, it sets the 
blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
         ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
         r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                  ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                  ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);
Top 21 bits are taken as it is. However, only 9 bit should be taken. So, 
for random values, it corrupts the startoff which from 9-63 bits.

         r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                  ((xfs_bmbt_rec_base_t)s->br_blockcount &
                  (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being 
done here. I would appreciate if someone can verify that assertion is 
correct.


Regards,
Shailendra