Re: [PATCH] workqueue: Add WQ_SCHED_FIFO

From: Nathan Huckleberry <nhuck@google.com>
To: Tejun Heo <tj@kernel.org>
Cc: Sandeep Dhavale <dhavale@google.com>,
	Daeho Jeong <daehojeong@google.com>,
	Eric Biggers <ebiggers@kernel.org>,
	Sami Tolvanen <samitolvanen@google.com>,
	Lai Jiangshan <jiangshanlai@gmail.com>,
	Jonathan Corbet <corbet@lwn.net>,
	linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] workqueue: Add WQ_SCHED_FIFO
Date: Sat, 14 Jan 2023 13:00:00 -0800	[thread overview]
Message-ID: <CAJkfWY7vmvrU8sW3OWpSa9zygY=6e8BTTkktPe-VScdOcLL-sw@mail.gmail.com> (raw)
In-Reply-To: <Y8HI+42TxxlJxT6D@slm.duckdns.org>

On Fri, Jan 13, 2023 at 1:11 PM Tejun Heo <tj@kernel.org> wrote:
>
> On Fri, Jan 13, 2023 at 01:07:02PM -0800, Nathan Huckleberry wrote:
> > Add a WQ flag that allows workqueues to use SCHED_FIFO with the least
> > imporant RT priority.  This can reduce scheduler latency for IO
> > post-processing when the CPU is under load without impacting other RT
> > workloads.  This has been shown to improve app startup time on Android
> > [1].
> >
> > Scheduler latency affects several drivers as evidenced by [1], [2], [3],
> > [4].  Some of these drivers have moved post-processing into IRQ context.
> > However, this can cause latency spikes for real-time threads and jitter
> > related jank on Android.  Using a workqueue with SCHED_FIFO improves
> > scheduler latency without causing latency problems for RT threads.
> >
> > [1]:
> > https://lore.kernel.org/linux-erofs/20230106073502.4017276-1-dhavale@google.com/
> > [2]:
> > https://lore.kernel.org/linux-f2fs-devel/20220802192437.1895492-1-daeho43@gmail.com/
> > [3]:
> > https://lore.kernel.org/dm-devel/20220722093823.4158756-4-nhuck@google.com/
> > [4]:
> > https://lore.kernel.org/dm-crypt/20200706173731.3734-1-ignat@cloudflare.com/
> >
> > This change has been tested on dm-verity with the following fio config:
> >
> > [global]
> > time_based
> > runtime=120
> >
> > [do-verify]
> > ioengine=sync
> > filename=/dev/testing
> > rw=randread
> > direct=1
> >
> > [burn_8x90%_qsort]
> > ioengine=cpuio
> > cpuload=90
> > numjobs=8
> > cpumode=qsort
> >
> > Before:
> > clat (usec): min=13, max=23882, avg=29.56, stdev=113.29 READ:
> > bw=122MiB/s (128MB/s), 122MiB/s-122MiB/s (128MB/s-128MB/s), io=14.3GiB
> > (15.3GB), run=120001-120001msec
> >
> > After:
> > clat (usec): min=13, max=23137, avg=19.96, stdev=105.71 READ:
> > bw=180MiB/s (189MB/s), 180MiB/s-180MiB/s (189MB/s-189MB/s), io=21.1GiB
> > (22.7GB), run=120012-120012msec
>
> Given that its use case mostly intersects with WQ_HIGHPRI, would it make
> more sense to add a switch to alter its behavior instead? I don't really
> like the idea of pushing the decision between WQ_HIGHPRI and WQ_SCHED_FIFO
> to each user.

This sounds fine to me. How do you feel about a config flag to change
the default WQ_HIGHPRI scheduler policy and a sysfs node to update the
policy per workqueue?

Thanks,
Huck

>
> Thanks.

>
> --
> tejun