f2fs recovery flow is relying on dnode block link list, it means fsynced
file recovery depends on previous dnode's persistence in the list, so
during fsync() we should wait on all regular inode's dnode writebacked
before issuing flush.

By this way, we can avoid dnode block list being broken by out-of-order
IO submission due to IO scheduler or driver.

Sheng Yong helps to do the test with this patch:

Target:/data (f2fs, -)
64MB / 32768KB / 4KB / 8

1 / PERSIST / Index

Base:
	SEQ-RD(MB/s)	SEQ-WR(MB/s)	RND-RD(IOPS)	RND-WR(IOPS)	Insert(TPS)	Update(TPS)	Delete(TPS)
1	867.82		204.15		41440.03	41370.54	680.8		1025.94		1031.08
2	871.87		205.87		41370.3		40275.2		791.14		1065.84		1101.7
3	866.52		205.69		41795.67	40596.16	694.69		1037.16		1031.48
Avg	868.7366667	205.2366667	41535.33333	40747.3		722.21		1042.98		1054.753333

After:
	SEQ-RD(MB/s)	SEQ-WR(MB/s)	RND-RD(IOPS)	RND-WR(IOPS)	Insert(TPS)	Update(TPS)	Delete(TPS)
1	798.81		202.5		41143		40613.87	602.71		838.08		913.83
2	805.79		206.47		40297.2		41291.46	604.44		840.75		924.27
3	814.83		206.17		41209.57	40453.62	602.85		834.66		927.91
Avg	806.4766667	205.0466667	40883.25667	40786.31667	603.3333333	837.83		922.0033333

Patched/Original:
	0.928332713	0.999074239	0.984300676	1.000957528	0.835398753	0.803303994	0.874141189

It looks like atomic write will suffer performance regression.

I suspect that the criminal is that we forcing to wait all dnode being in
storage cache before we issue PREFLUSH+FUA.

BTW, will commit ("f2fs: don't need to wait for node writes for atomic write")
cause the problem: we will lose data of last transaction after SPO, even if
atomic write return no error:

- atomic_open();
- write() P1, P2, P3;
- atomic_commit();
 - writeback data: P1, P2, P3;
 - writeback node: N1, N2, N3;  <--- If N1, N2 is not writebacked, N3 with fsync_mark is
writebacked, In SPOR, we won't find N3 since node chain is broken, turns out that losing
last transaction.
 - preflush + fua;
- power-cut

If we don't wait dnode writeback for atomic_write:

	SEQ-RD(MB/s)	SEQ-WR(MB/s)	RND-RD(IOPS)	RND-WR(IOPS)	Insert(TPS)	Update(TPS)	Delete(TPS)
1	779.91		206.03		41621.5		40333.16	716.9		1038.21		1034.85
2	848.51		204.35		40082.44	39486.17	791.83		1119.96		1083.77
3	772.12		206.27		41335.25	41599.65	723.29		1055.07		971.92
Avg	800.18		205.55		41013.06333	40472.99333	744.0066667	1071.08		1030.18

Patched/Original:
	0.92108464	1.001526693	0.987425886	0.993268102	1.030180511	1.026942031	0.976702294

SQLite's performance recovers.

Jaegeuk:
"Practically, I don't see db corruption becase of this. We can excuse to lose
the last transaction."

Finally, we decide to keep original implementation of atomic write interface
sematics that we don't wait all dnode writeback before preflush+fua submission.
Signed-off-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>