12aceb89b0bc eventfd: convert to f_op->read_iter() causes 6.7% performance regression

[Hao Xu <haoxu@xxxxxxxxxxxxxxxxx>]

Hi Jens,
We found that 12aceb89b0bc eventfd: convert to f_op->read_iter()
introduced 6.7% peformance regression in test suite will-it-scale

The test commands are:
./runtest.py eventfd1 10 thread 1
and
./runtest.py eventfd1 10 thread 128

the test is a simple while which is like:
while(time not over) {
  write(eventfd);
  read(eventfd);
}

By perf analysis, I saw that the main cost is new_sync_read().
Though I think this regression only happens when there is no
'schedule/sleep'(otherwise sleep occupies most time), I feel it
still necessary to report it.