On Wed, Aug 30, 2017 at 05:12:10PM +0300, Amir Goldstein wrote:

On Wed, Aug 30, 2017 at 4:46 PM, Christoph Hellwig [off-list ref] wrote:

quoted

Oops.

Yes, except for the case of calling xlog_state_release_iclog
ourselves we can't claim that the cache has been flushed.  In
fact I'm not even sure that is save, if ic_refcnt makes
xlog_state_release_iclog exit early, so we might have to
pass additional information around there.

With this change, for the multiple fsyncing tasks workload, the majority won't
be optimized, so the remaining code is only really optimizing the single
process doing seldom fsync case.

I wonder if it's worth keeping the optimization around at all?

I think it is.  And we should get back to optimizing the multi-threaded
cases soon using something like a sequence counter that checks if we
issued a log sync after finishing the data writeback.