Hi,
在 2024/03/26 13:22, Jinpu Wang 写道:
Hi Song, hi Kuai,
ping, Any comments?
Thx!
On Thu, Mar 7, 2024 at 1:08 PM Jack Wang <jinpu.wang@xxxxxxxxx> wrote:
From: Florian-Ewald Mueller <florian-ewald.mueller@xxxxxxxxx>
Replace md_thread's wait_event()/wake_up() related calls with their
simple swait~ variants to improve performance as well as memory and
stack footprint.
Use the IDLE state for the worker thread put to sleep instead of
misusing the INTERRUPTIBLE state combined with flushing signals
just for not contributing to system's cpu load-average stats.
Also check for sleeping thread before attempting its wake_up in
md_wakeup_thread() for avoiding unnecessary spinlock contention.
I think it'll be better to split this into a seperate patch.
And can you check if we just add wq_has_sleeper(), will there be
performance improvement?
With this patch (backported) on a kernel 6.1, the IOPS improved
by around 4% with raid1 and/or raid5, in high IO load scenarios.
Can you be more specifical about your test? because from what I know,
IO fast path doesn't involved with daemon thread, and I don't understand
yet where the 4% improvement is from.
Thanks,
Kuai
