os/bluestore: fix global commit latency

[xiexingguo]

"start" is used to calculate the global bluestore commit latency
and hence shall not be updated at each internal state enter/exit.

Otherwise the l_bluestore_commit_lat counter won't reflect the
real commit latency precisely.

Signed-off-by: xie xingguo xie.xingguo@zte.com.cn

[liewegas]

I think the idea was for this to log the latency in each state. Perhaps the perfcounter just needs to be renamed so that it doesn't imply it's more than just a single state?

[xiexingguo]

But from FileStore, the os_apply_latency is recording the latency of a transaction from FileStore::queue_transactions to FileStore::_finish_op.

I think BlueStore shall at least keep pace with FileStore(though it has identical os_commit_latency and os_apply_latency) by updating start of txc at constructing only. And then by doing:

void BlueStore::_txc_finish_kv(TransContext *txc)
logger->tinc(l_bluestore_commit_lat, ceph_clock_now(g_ceph_context) - txc->start);

l_bluestore_commit_lat will reflects the exact latency of the transaction between arriving at BlueStore and finished, which is the same as FileStore.

Otherwise start is a pure dup of last_stamp and hence can be eliminated.