Currently cluster allocation is always trying to find a victim chain (a
chian has most space), and this may lead to poor performance because of
discontiguous allocation in some scenarios.

Our test case is block size 4k, cluster size 1M and mount option with
localalloc=2048 (2G), since a gd is 32256M (about 31.5G) and a localalloc
window is only 2G, creating 50G file will result in 2G from gd0, 2G from
gd1, ...

One way to improve performance is enlarge localalloc window size (max
31104M), but this will make end user feel that about 30G is suddenly
"missing", and localalloc currently do not support steal, which means one
node cannot use another node's localalloc even it is not used in fact.  So
using the last gd to record the allocation and continues with the gd if it
has enough space for a localalloc window can make the allocation as more
contiguous as possible.

Our test result is below (evaluated in IOPS), which is using iometer
running in VM, dynamic vhd virtual disk stored in ocfs2.

IO model                Original   After   Improved(%)
16K60%Write100%Random     703       876     24.59%
8K90%Write100%Random      735       827     12.59%
4K100%Write100%Random     859       915      6.52%
4K100%Read100%Random     2092      2600     24.30%

Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Tested-by: Norton Zhu <norton.zhu@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>