Message-ID: <Y+DP6V9fZG7XPPGy@debian>
Date:   Mon, 6 Feb 2023 18:01:13 +0800
From:   Gao Xiang <xiang@...nel.org>
To:     Sandeep Dhavale <dhavale@...gle.com>
Cc:     Gao Xiang <xiang@...nel.org>, Chao Yu <chao@...nel.org>,
        Yue Hu <huyue2@...lpad.com>,
        Jeffle Xu <jefflexu@...ux.alibaba.com>,
        kernel-team@...roid.com, linux-erofs@...ts.ozlabs.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v4] erofs: replace erofs_unzipd workqueue with per-cpu
 threads

Hi Sandeep,

On Fri, Jan 06, 2023 at 07:35:01AM +0000, Sandeep Dhavale wrote:
> Using per-cpu thread pool we can reduce the scheduling latency compared
> to workqueue implementation. With this patch scheduling latency and
> variation is reduced as per-cpu threads are high priority kthread_workers.
> 
> The results were evaluated on arm64 Android devices running 5.10 kernel.
> 
> The table below shows resulting improvements of total scheduling latency
> for the same app launch benchmark runs with 50 iterations. Scheduling
> latency is the latency between when the task (workqueue kworker vs
> kthread_worker) became eligible to run to when it actually started
> running.
> +-------------------------+-----------+----------------+---------+
> |                         | workqueue | kthread_worker |  diff   |
> +-------------------------+-----------+----------------+---------+
> | Average (us)            |     15253 |           2914 | -80.89% |
> | Median (us)             |     14001 |           2912 | -79.20% |
> | Minimum (us)            |      3117 |           1027 | -67.05% |
> | Maximum (us)            |     30170 |           3805 | -87.39% |
> | Standard deviation (us) |      7166 |            359 |         |
> +-------------------------+-----------+----------------+---------+
> 
> Background: Boot times and cold app launch benchmarks are very
> important to the android ecosystem as they directly translate to
> responsiveness from user point of view. While erofs provides
> a lot of important features like space savings, we saw some
> performance penalty in cold app launch benchmarks in few scenarios.
> Analysis showed that the significant variance was coming from the
> scheduling cost while decompression cost was more or less the same.
> 
> Having per-cpu thread pool we can see from the above table that this
> variation is reduced by ~80% on average. This problem was discussed
> at LPC 2022. Link to LPC 2022 slides and
> talk at [1]
> 
> [1] https://lpc.events/event/16/contributions/1338/
> 
> Signed-off-by: Sandeep Dhavale <dhavale@...gle.com>
> ---
> V3 -> V4
> * Updated commit message with background information
> V2 -> V3
> * Fix a warning Reported-by: kernel test robot <lkp@...el.com>
> V1 -> V2
> * Changed name of kthread_workers from z_erofs to erofs_worker
> * Added kernel configuration to run kthread_workers at normal or
>   high priority
> * Added cpu hotplug support
> * Added wrapped kthread_workers under worker_pool
> * Added one unbound thread in a pool to handle a context where
>   we already stopped per-cpu kthread worker
> * Updated commit message

I've just modified your v4 patch based on erofs -dev branch with
my previous suggestion [1], but I haven't tested it.

Could you help check if the updated patch looks good to you and
test it on your side?  If there are unexpected behaviors, please
help update as well, thanks!

[1] https://lore.kernel.org/r/5e1b7191-9ea6-3781-7928-72ac4cd88591@linux.alibaba.com/

Thanks,
Gao Xiang

>From 2e87235abc745c0fef8e32abcd3a51546b4378ad Mon Sep 17 00:00:00 2001
From: Sandeep Dhavale <dhavale@...gle.com>
Date: Mon, 6 Feb 2023 17:53:39 +0800
Subject: [PATCH] erofs: add per-cpu threads for decompression

Using per-cpu thread pool we can reduce the scheduling latency compared
to workqueue implementation. With this patch scheduling latency and
variation is reduced as per-cpu threads are high priority kthread_workers.

The results were evaluated on arm64 Android devices running 5.10 kernel.

The table below shows resulting improvements of total scheduling latency
for the same app launch benchmark runs with 50 iterations. Scheduling
latency is the latency between when the task (workqueue kworker vs
kthread_worker) became eligible to run to when it actually started
running.
+-------------------------+-----------+----------------+---------+
|                         | workqueue | kthread_worker |  diff   |
+-------------------------+-----------+----------------+---------+
| Average (us)            |     15253 |           2914 | -80.89% |
| Median (us)             |     14001 |           2912 | -79.20% |
| Minimum (us)            |      3117 |           1027 | -67.05% |
| Maximum (us)            |     30170 |           3805 | -87.39% |
| Standard deviation (us) |      7166 |            359 |         |
+-------------------------+-----------+----------------+---------+

Background: Boot times and cold app launch benchmarks are very
important to the android ecosystem as they directly translate to
responsiveness from user point of view. While erofs provides
a lot of important features like space savings, we saw some
performance penalty in cold app launch benchmarks in few scenarios.
Analysis showed that the significant variance was coming from the
scheduling cost while decompression cost was more or less the same.

Having per-cpu thread pool we can see from the above table that this
variation is reduced by ~80% on average. This problem was discussed
at LPC 2022. Link to LPC 2022 slides and
talk at [1]

[1] https://lpc.events/event/16/contributions/1338/

Signed-off-by: Sandeep Dhavale <dhavale@...gle.com>
Signed-off-by: Gao Xiang <hsiangkao@...ux.alibaba.com>
---
 fs/erofs/Kconfig |  18 +++++
 fs/erofs/zdata.c | 190 ++++++++++++++++++++++++++++++++++++++++++-----
 2 files changed, 189 insertions(+), 19 deletions(-)

diff --git a/fs/erofs/Kconfig b/fs/erofs/Kconfig
index 85490370e0ca..704fb59577e0 100644
--- a/fs/erofs/Kconfig
+++ b/fs/erofs/Kconfig
@@ -108,3 +108,21 @@ config EROFS_FS_ONDEMAND
 	  read support.
 
 	  If unsure, say N.
+
+config EROFS_FS_PCPU_KTHREAD
+	bool "EROFS per-cpu decompression kthread workers"
+	depends on EROFS_FS_ZIP
+	help
+	  Saying Y here enables per-CPU kthread workers pool to carry out
+	  async decompression for low latencies on some architectures.
+
+	  If unsure, say N.
+
+config EROFS_FS_PCPU_KTHREAD_HIPRI
+	bool "EROFS high priority per-CPU kthread workers"
+	depends on EROFS_FS_ZIP && EROFS_FS_PCPU_KTHREAD
+	help
+	  This permits EROFS to configure per-CPU kthread workers to run
+	  at higher priority.
+
+	  If unsure, say N.
diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
index 384f64292f73..73198f494a6a 100644
--- a/fs/erofs/zdata.c
+++ b/fs/erofs/zdata.c
@@ -7,6 +7,8 @@
 #include "compress.h"
 #include <linux/prefetch.h>
 #include <linux/psi.h>
+#include <linux/slab.h>
+#include <linux/cpuhotplug.h>
 
 #include <trace/events/erofs.h>
 
@@ -109,6 +111,7 @@ struct z_erofs_decompressqueue {
 	union {
 		struct completion done;
 		struct work_struct work;
+		struct kthread_work kthread_work;
 	} u;
 	bool eio, sync;
 };
@@ -341,24 +344,128 @@ static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
 
 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 
-void z_erofs_exit_zip_subsystem(void)
+#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
+static struct kthread_worker __rcu **z_erofs_pcpu_workers;
+
+static void erofs_destroy_percpu_workers(void)
 {
-	destroy_workqueue(z_erofs_workqueue);
-	z_erofs_destroy_pcluster_pool();
+	struct kthread_worker *worker;
+	unsigned int cpu;
+
+	for_each_possible_cpu(cpu) {
+		worker = rcu_dereference_protected(
+					z_erofs_pcpu_workers[cpu], 1);
+		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
+		if (worker)
+			kthread_destroy_worker(worker);
+	}
+	kfree(z_erofs_pcpu_workers);
 }
 
-static inline int z_erofs_init_workqueue(void)
+static struct kthread_worker *erofs_init_percpu_worker(int cpu)
 {
-	const unsigned int onlinecpus = num_possible_cpus();
+	struct kthread_worker *worker =
+		kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);
 
-	/*
-	 * no need to spawn too many threads, limiting threads could minimum
-	 * scheduling overhead, perhaps per-CPU threads should be better?
-	 */
-	z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
-					    WQ_UNBOUND | WQ_HIGHPRI,
-					    onlinecpus + onlinecpus / 4);
-	return z_erofs_workqueue ? 0 : -ENOMEM;
+	if (IS_ERR(worker))
+		return worker;
+	if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))
+		sched_set_fifo_low(worker->task);
+	else
+		sched_set_normal(worker->task, 0);
+	return worker;
+}
+
+static int erofs_init_percpu_workers(void)
+{
+	struct kthread_worker *worker;
+	unsigned int cpu;
+
+	z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),
+			sizeof(struct kthread_worker *), GFP_ATOMIC);
+	if (!z_erofs_pcpu_workers)
+		return -ENOMEM;
+
+	for_each_online_cpu(cpu) {	/* could miss cpu{off,on}line? */
+		worker = erofs_init_percpu_worker(cpu);
+		if (!IS_ERR(worker))
+			rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
+	}
+	return 0;
+}
+#else
+static inline void erofs_destroy_percpu_workers(void) {}
+static inline int erofs_init_percpu_workers(void) { return 0; }
+#endif
+
+#if defined(CONFIG_HOTPLUG_CPU) && defined(EROFS_FS_PCPU_KTHREAD)
+static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
+static enum cpuhp_state erofs_cpuhp_state;
+
+static int erofs_cpu_online(unsigned int cpu)
+{
+	struct kthread_worker *worker, *old;
+
+	worker = erofs_init_percpu_worker(cpu);
+	if (IS_ERR(worker))
+		return ERR_PTR(worker);
+
+	spin_lock(&z_erofs_pcpu_worker_lock);
+	old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
+			lockdep_is_held(&z_erofs_pcpu_worker_lock));
+	if (!old)
+		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
+	spin_unlock(&z_erofs_pcpu_worker_lock);
+	if (old)
+		kthread_destroy_worker(worker);
+	return 0;
+}
+
+static int erofs_cpu_offline(unsigned int cpu)
+{
+	struct kthread_worker *worker;
+
+	spin_lock(&z_erofs_pcpu_worker_lock);
+	worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
+			lockdep_is_held(&z_erofs_pcpu_worker_lock));
+	rcu_assign_pointer(worker_pool.workers[cpu], NULL);
+	spin_unlock(&z_erofs_pcpu_worker_lock);
+
+	synchronize_rcu();
+	if (worker)
+		kthread_destroy_worker(worker);
+	return 0;
+}
+
+static int erofs_cpu_hotplug_init(void)
+{
+	int state;
+
+	state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
+			"fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);
+	if (state < 0)
+		return state;
+
+	erofs_cpuhp_state = state;
+	return 0;
+}
+
+static void erofs_cpu_hotplug_destroy(void)
+{
+	if (erofs_cpuhp_state)
+		cpuhp_remove_state_nocalls(erofs_cpuhp_state);
+}
+#else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */
+static inline int erofs_cpu_hotplug_init(void) { return 0; }
+static inline void erofs_cpu_hotplug_destroy(void) {}
+#endif
+
+void z_erofs_exit_zip_subsystem(void)
+{
+	erofs_cpu_hotplug_destroy();
+	erofs_destroy_percpu_workers();
+	destroy_workqueue(z_erofs_workqueue);
+	z_erofs_destroy_pcluster_pool();
 }
 
 int __init z_erofs_init_zip_subsystem(void)
@@ -366,10 +473,29 @@ int __init z_erofs_init_zip_subsystem(void)
 	int err = z_erofs_create_pcluster_pool();
 
 	if (err)
-		return err;
-	err = z_erofs_init_workqueue();
+		goto out_error_pcluster_pool;
+
+	z_erofs_workqueue = alloc_workqueue("erofs_worker",
+			WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());
+	if (!z_erofs_workqueue)
+		goto out_error_workqueue_init;
+
+	err = erofs_init_percpu_workers();
 	if (err)
-		z_erofs_destroy_pcluster_pool();
+		goto out_error_pcpu_worker;
+
+	err = erofs_cpu_hotplug_init();
+	if (err < 0)
+		goto out_error_cpuhp_init;
+	return err;
+
+out_error_cpuhp_init:
+	erofs_destroy_percpu_workers();
+out_error_pcpu_worker:
+	destroy_workqueue(z_erofs_workqueue);
+out_error_workqueue_init:
+	z_erofs_destroy_pcluster_pool();
+out_error_pcluster_pool:
 	return err;
 }
 
@@ -1305,11 +1431,17 @@ static void z_erofs_decompressqueue_work(struct work_struct *work)
 
 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 	z_erofs_decompress_queue(bgq, &pagepool);
-
 	erofs_release_pages(&pagepool);
 	kvfree(bgq);
 }
 
+#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
+static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
+{
+	z_erofs_decompressqueue_work((struct work_struct *)work);
+}
+#endif
+
 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 				       int bios)
 {
@@ -1324,9 +1456,24 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 
 	if (atomic_add_return(bios, &io->pending_bios))
 		return;
-	/* Use workqueue and sync decompression for atomic contexts only */
+	/* Use (kthread_)work and sync decompression for atomic contexts only */
 	if (in_atomic() || irqs_disabled()) {
+#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
+		struct kthread_worker *worker;
+
+		rcu_read_lock();
+		worker = rcu_dereference(
+				z_erofs_pcpu_workers[raw_smp_processor_id()]);
+		if (!worker) {
+			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
+			queue_work(z_erofs_workqueue, &io->u.work);
+		} else {
+			kthread_queue_work(worker, &io->u.kthread_work);
+		}
+		rcu_read_unlock();
+#else
 		queue_work(z_erofs_workqueue, &io->u.work);
+#endif
 		/* enable sync decompression for readahead */
 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
@@ -1455,7 +1602,12 @@ static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
 			*fg = true;
 			goto fg_out;
 		}
+#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
+		kthread_init_work(&q->u.kthread_work,
+				  z_erofs_decompressqueue_kthread_work);
+#else
 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
+#endif
 	} else {
 fg_out:
 		q = fgq;
@@ -1640,7 +1792,7 @@ static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
 
 	/*
 	 * although background is preferred, no one is pending for submission.
-	 * don't issue workqueue for decompression but drop it directly instead.
+	 * don't issue decompression but drop it directly instead.
 	 */
 	if (!*force_fg && !nr_bios) {
 		kvfree(q[JQ_SUBMIT]);
-- 
2.30.2

Powered by blists - more mailing lists