On Fri, May 28, 2021 at 4:56 PM Greg Kroah-Hartman
[off-list ref] wrote:

Fox, can you take some time and test these to verify it all still works
properly with your benchmarks?

I've tested it on an AWS C5a (amd, 96 logical cores):
Before, mutex_locks in kernfs_iop_permission(), kernfs_dop_revalidate
take significant time.
With the patchset, there is no mutex_lock issue. (see flamegraph
before.png/after.png)

On AWS C5 (intel, also 96 logical cores), the benchmark runs slower
than on c5a. But I don't think it's related, because
running the benchmark on ext4 is slower too, and the perf report,
which is no different from running on kernfs with this patchset,
shows the pressure is on the VFS side.

My conclusion: It works well with my benchmark.

I've attached:
flame graphs -- before.png/after.png
benchmark outputs -- result.before/result.after
perf reports -- report.before/report.after
perf report on ext4 -- report.baremetal.ext4
for you reference.


thanks,
fox