Re: [fsnotify] 4c40d6efc8: unixbench.score -3.3% regression

On Sat, Jul 31, 2021 at 10:51 PM Gabriel Krisman Bertazi
[off-list ref] wrote:

Amir Goldstein [off-list ref] writes:

quoted

On Sat, Jul 31, 2021 at 9:20 AM kernel test robot [off-list ref] wrote:

quoted

Greeting,

FYI, we noticed a -3.3% regression of unixbench.score due to commit:


commit: 4c40d6efc8b22b88a45c335ffd6d25b55d769f5b ("[PATCH v4 08/16] fsnotify: pass arguments of fsnotify() in struct fsnotify_event_info")
url: https://github.com/0day-ci/linux/commits/Gabriel-Krisman-Bertazi/File-system-wide-monitoring/20210721-001444
base: https://git.kernel.org/cgit/linux/kernel/git/jack/linux-fs.git fsnotify

in testcase: unixbench
on test machine: 96 threads 2 sockets Intel(R) Xeon(R) CPU @ 2.30GHz with 128G memory
with following parameters:

        runtime: 300s
        nr_task: 1
        test: pipe
        cpufreq_governor: performance
        ucode: 0x4003006

test-description: UnixBench is the original BYTE UNIX benchmark suite aims to test performance of Unix-like system.
test-url: https://github.com/kdlucas/byte-unixbench

In addition to that, the commit also has significant impact on the following tests:

+------------------+-------------------------------------------------------------------------------------+
| testcase: change | will-it-scale: will-it-scale.per_thread_ops -1.3% regression                        |
| test machine     | 192 threads 4 sockets Intel(R) Xeon(R) Platinum 9242 CPU @ 2.30GHz with 192G memory |
| test parameters  | cpufreq_governor=performance                                                        |
|                  | mode=thread                                                                         |
|                  | nr_task=100%                                                                        |
|                  | test=eventfd1                                                                       |
|                  | ucode=0x5003006                                                                     |
+------------------+-------------------------------------------------------------------------------------+


If you fix the issue, kindly add following tag
Reported-by: kernel test robot <redacted>

Gabriel,

It looks like my change throws away much of the performance gain for
small IO on pipes without any watches that was achieved by commit
71d734103edf ("fsnotify: Rearrange fast path to minimise overhead
when there is no watcher").

I think the way to fix it is to lift the optimization in __fsnotify()
to the fsnotify_parent() inline wrapper as Mel considered doing
but was not sure it was worth the effort at the time.

It's not completely trivial. I think it requires setting a flag
MNT_FSNOTIFY_WATCHED when there are watches on the
vfsmount. I will look into it.

Amir,

Since this patch is a clean up, would you mind if I drop it from my
series and base my work on top of mainline? Eventually, we can rebase
this patch, when the performance issue is addressed.

I ask because I'm about to send a v5 and I'm not sure if I should wait
to have this fixed.

I guess you mean that you want to add the sb to fsnotify() args list.
I don't mind, it's up to Jan.

Thanks,
Amir.