On 10/16, Christoph Hellwig wrote:
>The only change for the non-nowait case is that we now do a trylock before
>locking i_rwsem. In the past that was the more optimal pattern. Can you
>test the patch below if that's not the case anymore? We have a few more
>instances like that which might also want to be changed then.
>
The test result showed this fix patch helps recover the performance.
Below is the comparison (41d4b81 is the fix commit applied on top of
91f9943e1c).
$ compare -at 91f9943e1c7b6638f27312d03f
41d4b81c40b8873aa29b4b0224da3caf0242dd1c
tests: 1
testcase/path_params/tbox_group/run:
aim7/4BRD_12G-RAID1-xfs-disk_rd-9000-performance/lkp-ivb-ep01
91f9943e1c7b6638 41d4b81c40b8873aa29b4b0224
---------------- --------------------------
%stddev change %stddev
\ | \
353690 32% 467669 aim7.jobs-per-min
153 -24% 116 aim7.time.elapsed_time
153 -24% 116 aim7.time.elapsed_time.max
5670 -26% 4184 ± 3% aim7.time.system_time
69648 ± 15% -30% 48932 ± 11%
aim7.time.voluntary_context_switches
774279 ± 4% -33% 522402 ± 6%
aim7.time.involuntary_context_switches
63750 -20% 51085
interrupts.CAL:Function_call_interrupts
43353 4% 45157 vmstat.system.in
Thanks,
Xiaolong
>diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
>index 309e26c9dddb..ad61c78605d1 100644
>--- a/fs/xfs/xfs_file.c
>+++ b/fs/xfs/xfs_file.c
>@@ -258,10 +258,11 @@ xfs_file_buffered_aio_read(
> ssize_t ret;
>
> trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
>-
>- if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
>- if (iocb->ki_flags & IOCB_NOWAIT)
>+
>+ if (iocb->ki_flags & IOCB_NOWAIT) {
>+ if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
> return -EAGAIN;
>+ } else {
> xfs_ilock(ip, XFS_IOLOCK_SHARED);
> }
> ret = generic_file_read_iter(iocb, to);
