Skip to contentSkip to conversation

UFS Clock Scaling fixes and enhancements

PATCH Applied Asutosh Das 4 participants 9 messages Mar 25, 2020 – May 28, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v2 cover UFS Clock Scaling fixes and enhancements

Enhancements to clock-scaling parameters.
Few bug fixes to clock-scaling.

Changes since v1:
- Addressed comments of v1

Asutosh Das (3):
scsi: ufshcd: Update the set frequency to devfreq
scsi: ufshcd: Let vendor override devfreq parameters
scsi: ufs-qcom: Override devfreq parameters

 drivers/scsi/ufs/ufs-qcom.c | 25 +++++++++++++++++++++++++
 drivers/scsi/ufs/ufshcd.c   | 32 ++++++++++++++++++++++++++++++--
 drivers/scsi/ufs/ufshcd.h   | 12 ++++++++++++
 3 files changed, 67 insertions(+), 2 deletions(-)
Signature · 1 line
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
#1 lore raw

PATCH v2 1/3 scsi: ufshcd: Update the set frequency to devfreq

Currently, the frequency that devfreq provides the driver to set always leads the clocks to be scaled up. Hence, round the clock-rate to the nearest frequency before deciding to scale.

Also update the devfreq statistics of current frequency.

---
 drivers/scsi/ufs/ufshcd.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +16 -1
index 2a2a63b..7916c8b 100644
@@ -1187,6 +1187,9 @@ static int ufshcd_devfreq_target(struct device *dev,
 	if (!ufshcd_is_clkscaling_supported(hba))
 		return -EINVAL;
 
+	clki = list_first_entry(&hba->clk_list_head, struct ufs_clk_info, list);
+	/* Override with the closest supported frequency */
+	*freq = (unsigned long) clk_round_rate(clki->clk, *freq);
 	spin_lock_irqsave(hba->host->host_lock, irq_flags);
 	if (ufshcd_eh_in_progress(hba)) {
 		spin_unlock_irqrestore(hba->host->host_lock, irq_flags);
@@ -1201,8 +1204,11 @@ static int ufshcd_devfreq_target(struct device *dev,
 		goto out;
 	}
 
-	clki = list_first_entry(&hba->clk_list_head, struct ufs_clk_info, list);
+	/* Decide based on the rounded-off frequency and update */
 	scale_up = (*freq == clki->max_freq) ? true : false;
+	if (!scale_up)
+		*freq = clki->min_freq;
+	/* Update the frequency */
 	if (!ufshcd_is_devfreq_scaling_required(hba, scale_up)) {
 		spin_unlock_irqrestore(hba->host->host_lock, irq_flags);
 		ret = 0;
@@ -1250,6 +1256,8 @@ static int ufshcd_devfreq_get_dev_status(struct device *dev,
 	struct ufs_hba *hba = dev_get_drvdata(dev);
 	struct ufs_clk_scaling *scaling = &hba->clk_scaling;
 	unsigned long flags;
+	struct list_head *clk_list = &hba->clk_list_head;
+	struct ufs_clk_info *clki;
 
 	if (!ufshcd_is_clkscaling_supported(hba))
 		return -EINVAL;
@@ -1260,6 +1268,13 @@ static int ufshcd_devfreq_get_dev_status(struct device *dev,
 	if (!scaling->window_start_t)
 		goto start_window;
 
+	clki = list_first_entry(clk_list, struct ufs_clk_info, list);
+	/*
+	 * If current frequency is 0, then the ondemand governor considers
+	 * there's no initial frequency set. And it always requests to set
+	 * to max. frequency.
+	 */
+	stat->current_frequency = clki->curr_freq;
 	if (scaling->is_busy_started)
 		scaling->tot_busy_t += ktime_to_us(ktime_sub(ktime_get(),
 					scaling->busy_start_t));
Signature · 1 line
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
#2 lore raw

Re: [PATCH v2 1/3] scsi: ufshcd: Update the set frequency to devfreq

Quoted 9 lines · Asutosh Das

Currently, the frequency that devfreq provides the driver to set always leads the clocks to be scaled up. Hence, round the clock-rate to the nearest frequency before deciding to scale.

Also update the devfreq statistics of current frequency.

This change appears to cause issues for the Lenovo Miix 630, as identified by git bisect.

On 5.6-final, My boot log looks normal. On 5.7-rc7, the Lenovo Miix 630 rarely boots, usually stuck in some kind of infinite printk loop.

If I disable some of the UFS logging, I can capture this from the logs, as soon as UFS inits -

[    4.353860] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.359605] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.365412] ufshcd-qcom 1da4000.ufshc: ufshcd_check_errors:
saved_err 0x4 saved_uic_err 0x2
[    4.371121] ufshcd-qcom 1da4000.ufshc: hba->ufs_version = 0x210,
hba->capabilities = 0x1587001f
[    4.376846] ufshcd-qcom 1da4000.ufshc: hba->outstanding_reqs =
0x100000, hba->outstanding_tasks = 0x0
[    4.382636] ufshcd-qcom 1da4000.ufshc: last_hibern8_exit_tstamp at
0 us, hibern8_exit_cnt = 0
[    4.388368] ufshcd-qcom 1da4000.ufshc: No record of pa_err
[    4.394001] ufshcd-qcom 1da4000.ufshc: dl_err[0] = 0x80000001 at 3873626 us
[    4.399577] ufshcd-qcom 1da4000.ufshc: No record of nl_err
[    4.405053] ufshcd-qcom 1da4000.ufshc: No record of tl_err
[    4.410464] ufshcd-qcom 1da4000.ufshc: No record of dme_err
[    4.415747] ufshcd-qcom 1da4000.ufshc: No record of auto_hibern8_err
[    4.420950] ufshcd-qcom 1da4000.ufshc: No record of fatal_err
[    4.426013] ufshcd-qcom 1da4000.ufshc: No record of link_startup_fail
[    4.430950] ufshcd-qcom 1da4000.ufshc: No record of resume_fail
[    4.435786] ufshcd-qcom 1da4000.ufshc: No record of suspend_fail
[    4.440538] ufshcd-qcom 1da4000.ufshc: dev_reset[0] = 0x0 at 3031009 us
[    4.445199] ufshcd-qcom 1da4000.ufshc: No record of host_reset
[    4.449750] ufshcd-qcom 1da4000.ufshc: No record of task_abort
[    4.454214] ufshcd-qcom 1da4000.ufshc: clk: core_clk, rate: 50000000
[    4.458590] ufshcd-qcom 1da4000.ufshc: clk: core_clk_unipro, rate: 37500000

I don't understand how this change is breaking things, but it clearly is for me.

What kind of additional data would be useful to get to the bottom of this?

#3 lore raw

Re: [PATCH v2 1/3] scsi: ufshcd: Update the set frequency to devfreq

Hi Jeffrey

Quoted 7 lines · Jeffrey Hugo
Quoted 6 lines · Asutosh Das

Currently, the frequency that devfreq provides the driver to set always leads the clocks to be scaled up. Hence, round the clock-rate to the nearest frequency before deciding to scale.

Quoted 7 lines
Quoted 3 lines

Also update the devfreq statistics of current frequency.

This change appears to cause issues for the Lenovo Miix 630, as identified by git bisect.

Thanks for reporting this.

Quoted 34 lines

On 5.6-final, My boot log looks normal. On 5.7-rc7, the Lenovo Miix 630 rarely boots, usually stuck in some kind of infinite printk loop.

If I disable some of the UFS logging, I can capture this from the logs, as soon as UFS inits -

[    4.353860] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.359605] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.365412] ufshcd-qcom 1da4000.ufshc: ufshcd_check_errors:
saved_err 0x4 saved_uic_err 0x2
[    4.371121] ufshcd-qcom 1da4000.ufshc: hba->ufs_version = 0x210,
hba->capabilities = 0x1587001f
[    4.376846] ufshcd-qcom 1da4000.ufshc: hba->outstanding_reqs =
0x100000, hba->outstanding_tasks = 0x0
[    4.382636] ufshcd-qcom 1da4000.ufshc: last_hibern8_exit_tstamp at
0 us, hibern8_exit_cnt = 0
[    4.388368] ufshcd-qcom 1da4000.ufshc: No record of pa_err
[    4.394001] ufshcd-qcom 1da4000.ufshc: dl_err[0] = 0x80000001 at 3873626 us
[    4.399577] ufshcd-qcom 1da4000.ufshc: No record of nl_err
[    4.405053] ufshcd-qcom 1da4000.ufshc: No record of tl_err
[    4.410464] ufshcd-qcom 1da4000.ufshc: No record of dme_err
[    4.415747] ufshcd-qcom 1da4000.ufshc: No record of auto_hibern8_err
[    4.420950] ufshcd-qcom 1da4000.ufshc: No record of fatal_err
[    4.426013] ufshcd-qcom 1da4000.ufshc: No record of link_startup_fail
[    4.430950] ufshcd-qcom 1da4000.ufshc: No record of resume_fail
[    4.435786] ufshcd-qcom 1da4000.ufshc: No record of suspend_fail
[    4.440538] ufshcd-qcom 1da4000.ufshc: dev_reset[0] = 0x0 at 3031009 us
[    4.445199] ufshcd-qcom 1da4000.ufshc: No record of host_reset
[    4.449750] ufshcd-qcom 1da4000.ufshc: No record of task_abort
[    4.454214] ufshcd-qcom 1da4000.ufshc: clk: core_clk, rate: 50000000
[    4.458590] ufshcd-qcom 1da4000.ufshc: clk: core_clk_unipro, rate: 37500000
Quoted 4 lines

I don't understand how this change is breaking things, but it clearly is for me.

What kind of additional data would be useful to get to the bottom of this?

++

Let me take a look and get back on this.

Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
Linux Foundation Collaborative Project
#4 lore raw

Re: [PATCH v2 1/3] scsi: ufshcd: Update the set frequency to devfreq

Quoted 12 lines · Asutosh Das (asd)

Hi Jeffrey

Quoted 9 lines · Jeffrey Hugo
Quoted 8 lines · Asutosh Das

Currently, the frequency that devfreq provides the driver to set always leads the clocks to be scaled up. Hence, round the clock-rate to the nearest frequency before deciding to scale.

Also update the devfreq statistics of current frequency.

Quoted 46 lines
Quoted 5 lines
Quoted 1 line

This change appears to cause issues for the Lenovo Miix 630, as identified by git bisect.

Thanks for reporting this.

Quoted 38 lines

On 5.6-final, My boot log looks normal. On 5.7-rc7, the Lenovo Miix 630 rarely boots, usually stuck in some kind of infinite printk loop.

If I disable some of the UFS logging, I can capture this from the logs, as soon as UFS inits -

[    4.353860] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.359605] ufshcd-qcom 1da4000.ufshc: ufshcd_intr: Unhandled
interrupt 0x00000000
[    4.365412] ufshcd-qcom 1da4000.ufshc: ufshcd_check_errors:
saved_err 0x4 saved_uic_err 0x2
[    4.371121] ufshcd-qcom 1da4000.ufshc: hba->ufs_version = 0x210,
hba->capabilities = 0x1587001f
[    4.376846] ufshcd-qcom 1da4000.ufshc: hba->outstanding_reqs =
0x100000, hba->outstanding_tasks = 0x0
[    4.382636] ufshcd-qcom 1da4000.ufshc: last_hibern8_exit_tstamp at
0 us, hibern8_exit_cnt = 0
[    4.388368] ufshcd-qcom 1da4000.ufshc: No record of pa_err
[    4.394001] ufshcd-qcom 1da4000.ufshc: dl_err[0] = 0x80000001 at 3873626 us
[    4.399577] ufshcd-qcom 1da4000.ufshc: No record of nl_err
[    4.405053] ufshcd-qcom 1da4000.ufshc: No record of tl_err
[    4.410464] ufshcd-qcom 1da4000.ufshc: No record of dme_err
[    4.415747] ufshcd-qcom 1da4000.ufshc: No record of auto_hibern8_err
[    4.420950] ufshcd-qcom 1da4000.ufshc: No record of fatal_err
[    4.426013] ufshcd-qcom 1da4000.ufshc: No record of link_startup_fail
[    4.430950] ufshcd-qcom 1da4000.ufshc: No record of resume_fail
[    4.435786] ufshcd-qcom 1da4000.ufshc: No record of suspend_fail
[    4.440538] ufshcd-qcom 1da4000.ufshc: dev_reset[0] = 0x0 at 3031009 us
[    4.445199] ufshcd-qcom 1da4000.ufshc: No record of host_reset
[    4.449750] ufshcd-qcom 1da4000.ufshc: No record of task_abort
[    4.454214] ufshcd-qcom 1da4000.ufshc: clk: core_clk, rate: 50000000
[    4.458590] ufshcd-qcom 1da4000.ufshc: clk: core_clk_unipro, rate: 37500000

I don't understand how this change is breaking things, but it clearly is for me.

What kind of additional data would be useful to get to the bottom of this?

It turns out that the unipro_core clock had no parent, and thus no ability to scale. Fixing that in GCC seems to have resolved this. I suspect the UFS clock scaling code attempted to scale the core clock, didn't check that the clock could change rates, and went along assuming the new rate was effective, thus putting the hardware into a bad state.

#5 lore raw

PATCH v2 2/3 scsi: ufshcd: Let vendor override devfreq parameters

In reply to Asutosh Das

Vendor drivers may have a need to update the polling interval and thresholds.
Provide a vops for vendor drivers to use.

---
 drivers/scsi/ufs/ufshcd.c | 15 ++++++++++++++-
 drivers/scsi/ufs/ufshcd.h | 12 ++++++++++++
 2 files changed, 26 insertions(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +14 -1
index 7916c8b..61e8974 100644
@@ -1303,6 +1303,17 @@ static struct devfreq_dev_profile ufs_devfreq_profile = {
 	.get_dev_status	= ufshcd_devfreq_get_dev_status,
 };
 
+#if IS_ENABLED(CONFIG_DEVFREQ_GOV_SIMPLE_ONDEMAND)
+static struct devfreq_simple_ondemand_data ufs_ondemand_data = {
+	.upthreshold = 70,
+	.downdifferential = 5,
+};
+
+static void *gov_data = &ufs_ondemand_data;
+#else
+static void *gov_data = NULL;
+#endif
+
 static int ufshcd_devfreq_init(struct ufs_hba *hba)
 {
 	struct list_head *clk_list = &hba->clk_list_head;
@@ -1318,10 +1329,12 @@ static int ufshcd_devfreq_init(struct ufs_hba *hba)
 	dev_pm_opp_add(hba->dev, clki->min_freq, 0);
 	dev_pm_opp_add(hba->dev, clki->max_freq, 0);
 
+	ufshcd_vops_config_scaling_param(hba, &ufs_devfreq_profile,
+					 gov_data);
 	devfreq = devfreq_add_device(hba->dev,
 			&ufs_devfreq_profile,
 			DEVFREQ_GOV_SIMPLE_ONDEMAND,
-			NULL);
+			gov_data);
 	if (IS_ERR(devfreq)) {
 		ret = PTR_ERR(devfreq);
 		dev_err(hba->dev, "Unable to register with devfreq %d\n", ret);
drivers/scsi/ufs/ufshcd.h +12 -0
index d45a044..e8e8ee2 100644
@@ -56,6 +56,7 @@
 #include <linux/completion.h>
 #include <linux/regulator/consumer.h>
 #include <linux/bitfield.h>
+#include <linux/devfreq.h>
 #include "unipro.h"
 
 #include <asm/irq.h>
@@ -327,6 +328,9 @@ struct ufs_hba_variant_ops {
 	void	(*dbg_register_dump)(struct ufs_hba *hba);
 	int	(*phy_initialization)(struct ufs_hba *);
 	void	(*device_reset)(struct ufs_hba *hba);
+	void	(*config_scaling_param)(struct ufs_hba *hba,
+					struct devfreq_dev_profile *profile,
+					void *data);
 };
 
 /* clock gating state  */
@@ -1083,6 +1087,14 @@ static inline void ufshcd_vops_device_reset(struct ufs_hba *hba)
 	}
 }
 
+static inline void ufshcd_vops_config_scaling_param(struct ufs_hba *hba,
+						    struct devfreq_dev_profile
+						    *profile, void *data)
+{
+	if (hba->vops && hba->vops->config_scaling_param)
+		hba->vops->config_scaling_param(hba, profile, data);
+}
+
 extern struct ufs_pm_lvl_states ufs_pm_lvl_states[];
 
 /*
Signature · 1 line
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
#6 lore raw

PATCH v2 3/3 scsi: ufs-qcom: Override devfreq parameters

In reply to Asutosh Das

Override devfreq parameters for power-performance
trade-off.

---
 drivers/scsi/ufs/ufs-qcom.c | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)
drivers/scsi/ufs/ufs-qcom.c +25 -0
index 6115ac6..19aa5c4 100644
@@ -10,6 +10,7 @@
 #include <linux/phy/phy.h>
 #include <linux/gpio/consumer.h>
 #include <linux/reset-controller.h>
+#include <linux/devfreq.h>
 
 #include "ufshcd.h"
 #include "ufshcd-pltfrm.h"
@@ -1689,6 +1690,29 @@ static void ufs_qcom_device_reset(struct ufs_hba *hba)
 	usleep_range(10, 15);
 }
 
+#if IS_ENABLED(CONFIG_DEVFREQ_GOV_SIMPLE_ONDEMAND)
+static void ufs_qcom_config_scaling_param(struct ufs_hba *hba,
+					  struct devfreq_dev_profile *p,
+					  void *data)
+{
+	static struct devfreq_simple_ondemand_data *d;
+
+	if (!data)
+		return;
+
+	d = (struct devfreq_simple_ondemand_data *)data;
+	p->polling_ms = 60;
+	d->upthreshold = 70;
+	d->downdifferential = 5;
+}
+#else
+static void ufs_qcom_config_scaling_param(struct ufs_hba *hba,
+					  struct devfreq_dev_profile *p,
+					  void *data)
+{
+}
+#endif
+
 /**
  * struct ufs_hba_qcom_vops - UFS QCOM specific variant operations
  *
@@ -1710,6 +1734,7 @@ static const struct ufs_hba_variant_ops ufs_hba_qcom_vops = {
 	.resume			= ufs_qcom_resume,
 	.dbg_register_dump	= ufs_qcom_dump_dbg_regs,
 	.device_reset		= ufs_qcom_device_reset,
+	.config_scaling_param = ufs_qcom_config_scaling_param,
 };
 
 /**
Signature · 1 line
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
#7 lore raw
In reply to Asutosh Das

Looks good to me.
Thanks,
Avri

Quoted 21 lines

Enhancements to clock-scaling parameters.
Few bug fixes to clock-scaling.

Changes since v1:
- Addressed comments of v1

Asutosh Das (3):
scsi: ufshcd: Update the set frequency to devfreq
scsi: ufshcd: Let vendor override devfreq parameters
scsi: ufs-qcom: Override devfreq parameters

 drivers/scsi/ufs/ufs-qcom.c | 25 +++++++++++++++++++++++++
 drivers/scsi/ufs/ufshcd.c   | 32 ++++++++++++++++++++++++++++++--
 drivers/scsi/ufs/ufshcd.h   | 12 ++++++++++++
 3 files changed, 67 insertions(+), 2 deletions(-)

--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.

#8 lore raw
In reply to Asutosh Das

Asutosh,

Quoted 2 lines

Enhancements to clock-scaling parameters. Few bug fixes to clock-scaling.

Applied to 5.7/scsi-queue, thanks!

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering