xfs: fix an agbno overflow in __xfs_getfsmap_datadev

Dave Chinner reported that xfs/273 fails if the AG size happens to be an
exact power of two.  I traced this to an agbno integer overflow when the
current GETFSMAP call is a continuation of a previous GETFSMAP call, and
the last record returned was non-shareable space at the end of an AG.

__xfs_getfsmap_datadev sets up a data device query by converting the
incoming fmr_physical into an xfs_fsblock_t and cracking it into an agno
and agbno pair.  In the (failing) case of where fmr_blockcount of the
low key is nonzero and the record was for a non-shareable extent, it
will add fmr_blockcount to start_fsb and info->low.rm_startblock.

If the low key was actually the last record for that AG, then this
addition causes info->low.rm_startblock to point beyond EOAG.  When the
rmapbt range query starts, it'll return an empty set, and fsmap moves on
to the next AG.

Or so I thought.  Remember how we added to start_fsb?

If agsize < 1<<agblklog, start_fsb points to the same AG as the original
fmr_physical from the low key.  We run the rmapbt query, which returns
nothing, so getfsmap zeroes info->low and moves on to the next AG.

If agsize == 1<<agblklog, start_fsb now points to the next AG.  We run
the rmapbt query on the next AG with the excessively large
rm_startblock.  If this next AG is actually the last AG, we'll set
info->high to EOFS (which is now has a lower rm_startblock than
info->low), and the ranged btree query code will return -EINVAL.  If
it's not the last AG, we ignore all records for the intermediate AGs.

Oops.

Fix this by decoding start_fsb into agno and agbno only after making
adjustments to start_fsb.  This means that info->low.rm_startblock will
always be set to a valid agbno, and we always start the rmapbt iteration
in the correct AG.

While we're at it, fix the predicate for determining if an fsmap record
represents non-shareable space to include file data on pre-reflink
filesystems.

Reported-by: Dave Chinner <david@fromorbit.com>
Fixes: 63ef7a35 ("xfs: fix interval filtering in multi-step fsmap queries")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>