[PATCH 0/3] block: device frequency PM QoS tuning

[PATCH 0/3] block: device frequency PM QoS tuning

[Wang Jianzheng]

Hello,

These patches intruduce a mechanism for limiting device frequencies via
PM QoS when latency-sensitive threads block on IO. Stroage device (like
UFS) use the "devfreq_monitor" mechanism to automatically scale
frequency based on IO workloads. However, the hysteresis in IO workload
detection, it will lead IO request to be processed at low frequency. 
 
Original devfreq_monitir frequency scaling timeline:
       |--- latency-intensive process working ------------
  |****+**+**|***+++*++*|*++++++++*|**+++++***|*++++++++*|
       |                           |- high load and scale up frequency
       |-------- low frequency ----|-- high frequency ---|
([+] have IO request   [*] nothing to do)

Now, the patches provided here intruduce a mechanism for the block layer
to add constraints to device frequecny through PM QoS framework, with
configurable sysfs knobs per block device. Doing following config in my
test system:

  /sys/block/sda/dev_freq_timeout_ms = 30

This constraints is removed if there is no block IO for 30ms.

Enhanced frequency scaling timeline:
       |--- latency-intensive process working ------------
  |****+**+**|***+++*++*|*++++++++*|**+++++***|*++++++++*|
       |- add device frequecy PM QoS constraints----------
             |- scale up frequency
       |-low-|------------ high frequency ---------------|

Here are my example system detail:
  - SoC: Qualcomm Snapdragon (1+3+4 core cluster)
  - Stroage: UFS 4.1
  - Fio Version: 3.9
  - Global fio config:
           --rw=randread --bs=64k --iodepth=1 \
           --numjobs=5 --time_based --runtime=10 \
           --ioengine=libaio --hipri --cpus_allowed=3
           (job1~5 startdelay = [0s, 20s, 40s, 60s, 80s])
  - Local fio config:
      -Test case 1:
           --rate=10ms
      -Test case 2:
           --rate=0ms

Runing the same fio test used above with enhanced frequency scaling
enabled/disabled, I get:

  Test case 1:
     enabled: 	clat (usec): min=141, max=872, avg=550
     disabled:	clat (usec): min=210, max=899, avg=635
  Test case 2:
     enabled: 	BW=388.6(MB/s)
     disabled:	BW=378.2(MB/s)

So the intermittent workloads test(case 1) show >10% latency
improvement. The continuous workloads test(case2) show about 5%
bandwidth improvement.This mechanism delivers greater performance gains
under intermittent workloads compared to continuous workloads scenarios.

Any thoughts about the patches and the approach taken?

Wang Jianzheng (3):
  block/genhd: add sysfs knobs for the device frequency PM QoS
  block: add support for device frequency PM QoS tuning
  scsi: ufs: core: Add support for frequency PM QoS tuning

 block/blk-mq.c            | 58 +++++++++++++++++++++++++++++++++++++++
 block/genhd.c             | 23 ++++++++++++++++
 drivers/ufs/core/ufshcd.c | 44 +++++++++++++++++++++++++++++
 include/linux/blkdev.h    | 11 ++++++++
 include/linux/pm_qos.h    |  6 ++++
 5 files changed, 142 insertions(+)

-- 
2.34.1

[PATCH 1/3] block/genhd: add sysfs knobs for the device frequency PM QoS

[Wang Jianzheng]

Add sysfs knobs for the following parameters:

    dev_freq_timeout_ms: for clearing up the device frequency limit when
                        latency-intensive block IO is complete

This can be used to prevent delayed responses to latency-sensitive block
I/O operations when storage device operate at low frequency. By
implementing the "dev_freq_timeout_ms", it automatically restores device
frequency constraints throug PM QoS.

Signed-off-by: Wang Jianzheng <wangjianzheng@vivo.com>
---
 block/genhd.c          | 23 +++++++++++++++++++++++
 include/linux/blkdev.h |  2 ++
 2 files changed, 25 insertions(+)

diff --git a/block/genhd.c b/block/genhd.c
index 9bbc38d12792..9462a81501a8 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -1159,6 +1159,27 @@ static ssize_t partscan_show(struct device *dev,
 	return sysfs_emit(buf, "%u\n", disk_has_partscan(dev_to_disk(dev)));
 }
 
+static ssize_t dev_freq_timeout_ms_show(struct device *dev,
+					 struct device_attribute *attr, char *buf)
+{
+	struct gendisk *disk = dev_to_disk(dev);
+
+	return sprintf(buf, "%d\n", disk->dev_freq_timeout);
+}
+
+static ssize_t dev_freq_timeout_ms_store(struct device *dev,
+					  struct device_attribute *attr,
+					  const char *buf, size_t count)
+{
+	struct gendisk *disk = dev_to_disk(dev);
+	int i;
+
+	if (count > 0 && !kstrtoint(buf, 10, &i))
+		disk->dev_freq_timeout = i;
+
+	return count;
+}
+
 static DEVICE_ATTR(range, 0444, disk_range_show, NULL);
 static DEVICE_ATTR(ext_range, 0444, disk_ext_range_show, NULL);
 static DEVICE_ATTR(removable, 0444, disk_removable_show, NULL);
@@ -1173,6 +1194,7 @@ static DEVICE_ATTR(inflight, 0444, part_inflight_show, NULL);
 static DEVICE_ATTR(badblocks, 0644, disk_badblocks_show, disk_badblocks_store);
 static DEVICE_ATTR(diskseq, 0444, diskseq_show, NULL);
 static DEVICE_ATTR(partscan, 0444, partscan_show, NULL);
+static DEVICE_ATTR_RW(dev_freq_timeout_ms);
 
 #ifdef CONFIG_FAIL_MAKE_REQUEST
 ssize_t part_fail_show(struct device *dev,
@@ -1224,6 +1246,7 @@ static struct attribute *disk_attrs[] = {
 	&dev_attr_events_poll_msecs.attr,
 	&dev_attr_diskseq.attr,
 	&dev_attr_partscan.attr,
+	&dev_attr_dev_freq_timeout_ms.attr,
 #ifdef CONFIG_FAIL_MAKE_REQUEST
 	&dev_attr_fail.attr,
 #endif
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index fe1797bbec42..950ad047dd81 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -213,6 +213,8 @@ struct gendisk {
 	u64 diskseq;
 	blk_mode_t open_mode;
 
+	int dev_freq_timeout;
+
 	/*
 	 * Independent sector access ranges. This is always NULL for
 	 * devices that do not have multiple independent access ranges.
-- 
2.34.1

[PATCH 2/3] block: add support for device frequency PM QoS tuning

[Wang Jianzheng]

This mechanism use the PM QoS framework to add device frequency
constraints during Block IO is running.  When critical processing
I/O requests that could block latency-sensitive threads, it
dynamically applies frequency restrictions. These constraints will
be removed through a timeout-based reset mechanism.

Signed-off-by: Wang Jianzheng <wangjianzheng@vivo.com>
---
 block/blk-mq.c         | 58 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/blkdev.h |  9 +++++++
 include/linux/pm_qos.h |  6 +++++
 3 files changed, 73 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index ba3a4b77f578..fcb4034287d3 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -11,6 +11,7 @@
 #include <linux/bio.h>
 #include <linux/blkdev.h>
 #include <linux/blk-integrity.h>
+#include <linux/blk-pm.h>
 #include <linux/kmemleak.h>
 #include <linux/mm.h>
 #include <linux/init.h>
@@ -28,6 +29,7 @@
 #include <linux/prefetch.h>
 #include <linux/blk-crypto.h>
 #include <linux/part_stat.h>
+#include <linux/pm_qos.h>
 #include <linux/sched/isolation.h>
 
 #include <trace/events/block.h>
@@ -3095,6 +3097,54 @@ static bool bio_unaligned(const struct bio *bio, struct request_queue *q)
 	return false;
 }
 
+#ifdef CONFIG_PM
+static void blk_mq_dev_frequency_work(struct work_struct *work)
+{
+	struct request_queue *q =
+			container_of(work, struct request_queue, dev_freq_work.work);
+	unsigned long timeout;
+	struct dev_pm_qos_request *qos = q->dev_freq_qos;
+
+	timeout = msecs_to_jiffies(q->disk->dev_freq_timeout);
+	if (!q || IS_ERR_OR_NULL(q->dev) || IS_ERR_OR_NULL(qos))
+		return;
+
+	if (q->pm_qos_status == PM_QOS_ACTIVE) {
+		q->pm_qos_status = PM_QOS_FREQ_SET;
+		dev_pm_qos_add_request(q->dev, qos, DEV_PM_QOS_MIN_FREQUENCY,
+				       FREQ_QOS_MAX_DEFAULT_VALUE);
+	} else {
+		if (time_after(jiffies, READ_ONCE(q->last_active) + timeout))
+			q->pm_qos_status = PM_QOS_FREQ_REMOV;
+	}
+
+	if (q->pm_qos_status == PM_QOS_FREQ_REMOV) {
+		dev_pm_qos_remove_request(qos);
+		q->pm_qos_status = PM_QOS_ACTIVE;
+	} else {
+		schedule_delayed_work(&q->dev_freq_work,
+				      q->last_active + timeout - jiffies);
+	}
+}
+
+static void blk_pm_qos_dev_freq_update(struct request_queue *q, struct bio *bio)
+{
+	if (IS_ERR_OR_NULL(q->dev) || !q->disk->dev_freq_timeout)
+		return;
+
+	if ((bio->bi_opf & REQ_SYNC || bio_op(bio) == REQ_OP_READ)) {
+		WRITE_ONCE(q->last_active, jiffies);
+		if (q->pm_qos_status == PM_QOS_ACTIVE)
+			schedule_delayed_work(&q->dev_freq_work, 0);
+	}
+}
+#else
+static void blk_pm_qos_dev_freq_update(struct request_queue *q, struct bio *bio)
+{
+	return;
+}
+#endif
+
 /**
  * blk_mq_submit_bio - Create and send a request to block device.
  * @bio: Bio pointer.
@@ -3161,6 +3211,8 @@ void blk_mq_submit_bio(struct bio *bio)
 		goto queue_exit;
 	}
 
+	blk_pm_qos_dev_freq_update(q, bio);
+
 	bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
 	if (!bio)
 		goto queue_exit;
@@ -4601,6 +4653,12 @@ int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
 
 	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
+#ifdef CONFIG_PM
+	INIT_DELAYED_WORK(&q->dev_freq_work, blk_mq_dev_frequency_work);
+	q->dev_freq_qos = kzalloc(sizeof(*q->dev_freq_qos), GFP_KERNEL);
+	if (IS_ERR_OR_NULL(q->dev_freq_qos))
+		goto err_hctxs;
+#endif
 	INIT_LIST_HEAD(&q->flush_list);
 	INIT_LIST_HEAD(&q->requeue_list);
 	spin_lock_init(&q->requeue_lock);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 950ad047dd81..ea6dfff6b0f5 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -26,6 +26,7 @@
 #include <linux/xarray.h>
 #include <linux/file.h>
 #include <linux/lockdep.h>
+#include <linux/pm_qos.h>
 
 struct module;
 struct request_queue;
@@ -522,6 +523,14 @@ struct request_queue {
 #ifdef CONFIG_PM
 	struct device		*dev;
 	enum rpm_status		rpm_status;
+	enum pm_qos_status 	pm_qos_status;
+	unsigned long		last_active;
+
+	/** @dev_freq_work: Work to handle dev frequency PM QoS limits. */
+	struct delayed_work	dev_freq_work;
+
+	/** @dev_freq_qos: PM QoS frequency limits for dev. */
+	struct dev_pm_qos_request  *dev_freq_qos;
 #endif
 
 	/*
diff --git a/include/linux/pm_qos.h b/include/linux/pm_qos.h
index 4a69d4af3ff8..e0d77ff65942 100644
--- a/include/linux/pm_qos.h
+++ b/include/linux/pm_qos.h
@@ -95,6 +95,12 @@ struct freq_qos_request {
 	struct freq_constraints *qos;
 };
 
+enum pm_qos_status {
+	PM_QOS_INVALID = -1,
+	PM_QOS_ACTIVE = 0,
+	PM_QOS_FREQ_SET,
+	PM_QOS_FREQ_REMOV,
+};
 
 enum dev_pm_qos_req_type {
 	DEV_PM_QOS_RESUME_LATENCY = 1,
-- 
2.34.1

[PATCH 3/3] scsi: ufs: core: Add support for frequency PM QoS tuning

[Wang Jianzheng]

During frequency scaling operations, the host can reads the frequency
constraint configured in the PM QoS framework and updates the target
frequency accordingly.

Signed-off-by: Wang Jianzheng <wangjianzheng@vivo.com>
---
 drivers/ufs/core/ufshcd.c | 44 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)

diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c
index 9a43102b2b21..1a9bcbc9dab0 100644
--- a/drivers/ufs/core/ufshcd.c
+++ b/drivers/ufs/core/ufshcd.c
@@ -1546,6 +1546,47 @@ static void ufshcd_clk_scaling_resume_work(struct work_struct *work)
 	devfreq_resume_device(hba->devfreq);
 }
 
+#ifdef CONFIG_PM
+static void ufshcd_pm_qos_freq_read_value(struct ufs_hba *hba,
+					   unsigned long *freq)
+{
+	struct Scsi_Host *shost = hba->host;
+	struct scsi_device *sdev;
+	struct request_queue *q;
+	struct ufs_clk_info *clki;
+	s32 min_qos_freq;
+
+	if (!shost)
+		return;
+
+	shost_for_each_device(sdev, shost) {
+		if (!sdev)
+			continue;
+
+		q = sdev->request_queue;
+		if (IS_ERR_OR_NULL(q))
+			continue;
+
+		if (q->disk && !IS_ERR_OR_NULL(q->dev)) {
+			if (q->disk->dev_freq_timeout) {
+				min_qos_freq = dev_pm_qos_read_value(q->dev,
+								     DEV_PM_QOS_MIN_FREQUENCY);
+				clki = list_first_entry(&hba->clk_list_head,
+							struct ufs_clk_info, list);
+				if (min_qos_freq > clki->max_freq)
+					*freq = clki->max_freq;
+			}
+		}
+	}
+}
+#else
+static void ufshcd_pm_qos_freq_read_value(struct ufs_hba *hba,
+					   unsigned long *freq)
+{
+	return;
+}
+#endif
+
 static int ufshcd_devfreq_target(struct device *dev,
 				unsigned long *freq, u32 flags)
 {
@@ -1559,6 +1600,9 @@ static int ufshcd_devfreq_target(struct device *dev,
 	if (!ufshcd_is_clkscaling_supported(hba))
 		return -EINVAL;
 
+	/* Select the frequency based on the DEV_PM_QOS_MIN_FREQUENCY */
+	ufshcd_pm_qos_freq_read_value(hba, freq);
+
 	if (hba->use_pm_opp) {
 		struct dev_pm_opp *opp;
 
-- 
2.34.1

Re: [PATCH 0/3] block: device frequency PM QoS tuning

[Bart Van Assche]

On 9/14/25 4:45 AM, Wang Jianzheng wrote:
> Now, the patches provided here intruduce a mechanism for the block layer
> to add constraints to device frequecny through PM QoS framework, with
> configurable sysfs knobs per block device. Doing following config in my
> test system:
> 
>    /sys/block/sda/dev_freq_timeout_ms = 30
> 
> This constraints is removed if there is no block IO for 30ms.

Why a new sysfs attribute? Is this attribute really necessary? How are
users expected to determine what value to write into that attribute to
achieve optimal results?

Thanks,

Bart.

Re: [PATCH 1/3] block/genhd: add sysfs knobs for the device frequency PM QoS

[Bart Van Assche]

On 9/14/25 4:45 AM, Wang Jianzheng wrote:
>   block/genhd.c          | 23 +++++++++++++++++++++++
>   include/linux/blkdev.h |  2 ++

If  a new sysfs attribute is added then an entry must be added in
Documentation/ABI/ that explains the purpose of the new attribute and
also how to use it.

> --- a/include/linux/blkdev.h
> +++ b/include/linux/blkdev.h
> @@ -213,6 +213,8 @@ struct gendisk {
>   	u64 diskseq;
>   	blk_mode_t open_mode;
>   
> +	int dev_freq_timeout;

The name `dev_freq_timeout` is not sufficient to guess the meaning of
the attribute. Please add a comment.

Thanks,

Bart.

Re: [PATCH 3/3] scsi: ufs: core: Add support for frequency PM QoS tuning

[Bart Van Assche]

On 9/14/25 4:45 AM, Wang Jianzheng wrote:
> +	if (!shost)
> +		return;

Please remove the above if-statement and make sure that the
initialization order guarantees that hba->host is set before
ufshcd_pm_qos_freq_read_value() can be called.

> +	shost_for_each_device(sdev, shost) {

Why a loop over all logical units? Isn't it sufficient to check the
WLUN?

> +		if (!sdev)
> +			continue;
> +
> +		q = sdev->request_queue;
> +		if (IS_ERR_OR_NULL(q))
> +			continue;
> +
> +		if (q->disk && !IS_ERR_OR_NULL(q->dev)) {

The above three if-statements are not necessary. Please remove these
three if-statements to improve code clarity.

Thanks,

Bart.

Re: [PATCH 2/3] block: add support for device frequency PM QoS tuning

[Bart Van Assche]

On 9/14/25 4:45 AM, Wang Jianzheng wrote:
> +#ifdef CONFIG_PM
> +static void blk_mq_dev_frequency_work(struct work_struct *work)
> +{
> +	struct request_queue *q =
> +			container_of(work, struct request_queue, dev_freq_work.work);
> +	unsigned long timeout;
> +	struct dev_pm_qos_request *qos = q->dev_freq_qos;
> +
> +	timeout = msecs_to_jiffies(q->disk->dev_freq_timeout);
> +	if (!q || IS_ERR_OR_NULL(q->dev) || IS_ERR_OR_NULL(qos))
> +		return;
> +
> +	if (q->pm_qos_status == PM_QOS_ACTIVE) {
> +		q->pm_qos_status = PM_QOS_FREQ_SET;
> +		dev_pm_qos_add_request(q->dev, qos, DEV_PM_QOS_MIN_FREQUENCY,
> +				       FREQ_QOS_MAX_DEFAULT_VALUE);
> +	} else {
> +		if (time_after(jiffies, READ_ONCE(q->last_active) + timeout))
> +			q->pm_qos_status = PM_QOS_FREQ_REMOV;
> +	}
> +
> +	if (q->pm_qos_status == PM_QOS_FREQ_REMOV) {
> +		dev_pm_qos_remove_request(qos);
> +		q->pm_qos_status = PM_QOS_ACTIVE;
> +	} else {
> +		schedule_delayed_work(&q->dev_freq_work,
> +				      q->last_active + timeout - jiffies);
> +	}
> +}

The above code is similar in nature to the activity detection by the
run-time power management (RPM) code. Why a new timer mechanism instead
of adding more code in the UFS driver RPM callbacks?

> @@ -3161,6 +3211,8 @@ void blk_mq_submit_bio(struct bio *bio)
>   		goto queue_exit;
>   	}
>   
> +	blk_pm_qos_dev_freq_update(q, bio);

Good luck with adding power-management code in the block layer hot path
... I'm not sure anyone will be enthusiast seeing code being added in
blk_mq_submit_bio().

Thanks,

Bart.