LU-11912 ofd: reduce LUSTRE_DATA_SEQ_MAX_WIDTH

LU-11912 ofd: reduce LUSTRE_DATA_SEQ_MAX_WIDTH

Reduce LUSTRE_DATA_SEQ_MAX_WIDTH from ~4B to ~32M
to limit the number of objects under /O/[seq]/d[0..31]
dir on OSTs.
This makes the directories stay optimial for ldiskfs,
to avoid going into the largedir/3-level htree territory.

Remove the hard-coded LUSTRE_DATA_SEQ_MAX_WIDTH checks
in ofd, make them check the seq->lcs_width which is
a tunable set to LUSTRE_DATA_SEQ_MAX_WIDTH by default,
allow the value up to IDIF_MAX_OID if a larger seq width
is needed.

Use the odbo->o_size in the OST_CREATE rpc reply on ofd,
to update osp with the current seq width setting.
osp then uses this seq width to determine when to rollover
to a new seq.

The seq will rollover when the seq width is exhausted,
the default is LUSTRE_DATA_SEQ_MAX_WIDTH.
For seq >= FID_SEQ_NORMAL objects, the upper limit of
seq width is OBIF_MAX_OID,
For IDIF/MDT0 objects, the upper limit is IDIF_MAX_OID.
The seq FID_SEQ_OST_MDT0 will change to a normal seq after the
rollover.

Fix osp_precreate_reserve when the last precreated is the end
of the seq and the osp_objs_precreated can not host all
the requested objects, the mdt thread would stuck:
it wakes up osp precreate thread in a loop for progress,
but osp thread will not try to do anything until the seq
is used up. This can be seen easier when seq->lcs_width is
set to a low number and try to create an overstripe with stripe
number bigger than seq->lcs_width.

Fix the precreate thread spinning when the precreate pool
is at the end of the seq, and is nearly empty.

Change the seq->lcs_width to 16384 for all tests in
test-framework.sh, except a few slow tests to avoid timeouts,
and some overstriping tests creating LOV_MAX_STRIPE_COUNT to
avoid overstriping creating less objects than expected,
when precreate pool is at the end of the seq, and there are
not enough objects.

Fix the problem where seq could still change after
replay_barrier. To achieve this, introduce new fail_loc
OBD_FAIL_OSP_FORCE_NEW_SEQ and force_new_seq/force_new_seq_all
to drain the objects in the precreate pool then rollover to a
new seq. This applies to a bunch of test suites heavily using
replay_barrier.

Change-Id: I2749c1004b7bf3197b691cc94527f90145bcdef8
Signed-off-by: Li Dongyang <dongyangli@ddn.com>
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/38424
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Sergey Cheremencev <scherementsev@ddn.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>