On 5/6/21 3:08 PM, Al Viro wrote: > On Thu, May 06, 2021 at 01:15:01PM -0600, Jens Axboe wrote: > >> Attached output of perf annotate <func> for that last run. > > Heh... I wonder if keeping the value of iocb_flags(file) in > struct file itself would have a visible effect... A quick hack to get rid of the init_sync_kiocb() in new_sync_write() and just eliminate the ki_flags read in eventfd_write(), as the test case is blocking. That brings us closer to the ->write() method, down 7% vs the previous 10%: Executed in 468.23 millis fish external usr time 95.09 millis 114.00 micros 94.98 millis sys time 372.98 millis 76.00 micros 372.90 millis Executed in 468.97 millis fish external usr time 91.05 millis 89.00 micros 90.96 millis sys time 377.92 millis 69.00 micros 377.85 millis -- Jens Axboe
