Skip to contentSkip to conversation

Fix various issue on CPU freq for Armada 37xx

PATCH In review Gregory CLEMENT 3 participants 8 messages Mar 8, 2019 – Apr 29, 2019

Subsystems (3)
lore
Share & tools
lore

Gregory CLEMENT

author maintainer
#0 lore raw

cover Fix various issue on CPU freq for Armada 37xx

Hello,

A few weeks ago an issue was reported on the cpufreq driver, in the meantime Christian Neubert got the same issue on a ambian kernel and proposed a patch to fix it.

His fix seemed to address the issue seen, and I only added more context about what was done.

Christian also reported an other issue about the way the frequency was calculated. This issue is visible only when the default frequency was low (600MHz). Based on this I proposed to no more use the CPU frequency as base for the calculation but on the parent clock frequency

Gregory

Christian Neubert (1):
cpufreq: armada-37xx: fix clock parenting

Gregory CLEMENT (1):
cpufreq: armada-37xx: fix frequency calculation for opp

 drivers/cpufreq/armada-37xx-cpufreq.c | 42 +++++++++++++++++++++++----
 1 file changed, 36 insertions(+), 6 deletions(-)
Signature · 1 line
2.20.1
author maintainer
#1 lore raw

PATCH 1/2 cpufreq: armada-37xx: fix clock parenting

From: Christian Neubert <[email protected]>

The clock parenting was not setup properly when DVFS was enabled. It was expected that the same clock source was used with and without DVFS which was not the case.

This patch fixes this issue, allowing to make the cpufreq support work when the CPU clock source are not the default ones.

Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
[gregory: extract from a larger patch, modify comments and commit log]
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/cpufreq/armada-37xx-cpufreq.c +17 -3
index 75491fc841a6..ad4463e4266e 100644
@@ -162,11 +162,25 @@ static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base,
 	}
 
 	/*
-	 * Set cpu clock source, for all the level we keep the same
-	 * clock source that the one already configured. For this one
-	 * we need to use the clock framework
+	 * Set CPU clock source, for all the level we keep the same
+	 * clock source that the one already configured with DVS
+	 * disabled. For this one we need to use the clock framework
 	 */
 	parent = clk_get_parent(clk);
+
+	/*
+	 * Unset parent clock to force the clock framework setting again
+	 * the clock parent
+	 */
+	clk_set_parent(clk, NULL);
+
+	/*
+	 * For the Armada 37xx CPU clocks, setting the parent will
+	 * actually configure the parent when DVFS is enabled. At
+	 * hardware level it will be a different register from the one
+	 * read when doing clk_get_parent that will be set with
+	 * clk_set_parent.
+	 */
 	clk_set_parent(clk, parent);
 }
 
Signature · 1 line
2.20.1
maintainer
#2 lore raw

Re: [PATCH 1/2] cpufreq: armada-37xx: fix clock parenting

Quoted 23 lines · Gregory CLEMENT

From: Christian Neubert <[email protected]>

The clock parenting was not setup properly when DVFS was enabled. It was expected that the same clock source was used with and without DVFS which was not the case.

This patch fixes this issue, allowing to make the cpufreq support work when the CPU clock source are not the default ones.

Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
[gregory: extract from a larger patch, modify comments and commit log]
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/cpufreq/armada-37xx-cpufreq.c +0 -0
index 75491fc841a6..ad4463e4266e 100644

drivers/cpufreq/armada-37xx-cpufreq.c static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base, · line 166

Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +2 -3
@@ -162,11 +162,25 @@ static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base,
 	}
 
 	/*
-	 * Set cpu clock source, for all the level we keep the same
-	 * clock source that the one already configured. For this one
-	 * we need to use the clock framework
+	 * Set CPU clock source, for all the level we keep the same
+	 * clock source that the one already configured with DVS

DVFS ?

drivers/cpufreq/armada-37xx-cpufreq.c static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base, · line 167

Quoted 1 line
drivers/cpufreq/armada-37xx-cpufreq.c +1 -0
+	 * disabled. For this one we need to use the clock framework

Please add a full-stop (.) at the end.

Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +6 -0
 	 */
 	parent = clk_get_parent(clk);
+
+	/*
+	 * Unset parent clock to force the clock framework setting again
+	 * the clock parent
+	 */
+	clk_set_parent(clk, NULL);

drivers/cpufreq/armada-37xx-cpufreq.c static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base, · line 185

Quoted 10 lines
drivers/cpufreq/armada-37xx-cpufreq.c +8 -0
+
+	/*
+	 * For the Armada 37xx CPU clocks, setting the parent will
+	 * actually configure the parent when DVFS is enabled. At
+	 * hardware level it will be a different register from the one
+	 * read when doing clk_get_parent that will be set with
+	 * clk_set_parent.
+	 */
 	clk_set_parent(clk, parent);
 }

Maybe this should be done right from the clock driver instead? As cpufreq may or maynot be enabled by default (Surely most of the people will always enable it, but I am just trying to find the right place for doing this). The way we are setting the clock parent isn't that great, and looks a bit hacky just because of the way clock framework is. Maybe doing it directly, without getting clock framework in between, from the clock driver may look sane ?

Signature · 1 line
viresh
author maintainer
#3 lore raw

Re: [PATCH 1/2] cpufreq: armada-37xx: fix clock parenting

Hi Viresh,

Quoted 1 line · Viresh Kumar

On 08-03-19, 17:47, Gregory CLEMENT wrote:

[...]

Quoted 16 lines
Quoted 16 lines
drivers/cpufreq/armada-37xx-cpufreq.c +14 -0
 	parent = clk_get_parent(clk);
+
+	/*
+	 * Unset parent clock to force the clock framework setting again
+	 * the clock parent
+	 */
+	clk_set_parent(clk, NULL);
+
+	/*
+	 * For the Armada 37xx CPU clocks, setting the parent will
+	 * actually configure the parent when DVFS is enabled. At
+	 * hardware level it will be a different register from the one
+	 * read when doing clk_get_parent that will be set with
+	 * clk_set_parent.
+	 */
 	clk_set_parent(clk, parent);
Quoted 8 lines
Quoted 1 line

}

Maybe this should be done right from the clock driver instead? As cpufreq may or maynot be enabled by default (Surely most of the people will always enable it, but I am just trying to find the right place for doing this). The way we are setting the clock parent isn't that great, and looks a bit hacky just because of the way clock framework is. Maybe doing it directly, without getting clock framework in between, from the clock driver may look sane ?

I've just sent a patch following your suggestion to the clock mailing but keeping you in CC too.

Gregory

Quoted 3 lines
Signature · 2 lines
-- 
viresh
Signature · 3 lines
Gregory Clement, Bootlin
Embedded Linux and Kernel engineering
http://bootlin.com
#4 lore raw

Re: [PATCH 1/2] cpufreq: armada-37xx: fix clock parenting

In reply to Gregory CLEMENT

Hi Christian, Gregory,

Quoted 54 lines

From: Christian Neubert <[email protected]>

The clock parenting was not setup properly when DVFS was enabled. It was expected that the same clock source was used with and without DVFS which was not the case.

This patch fixes this issue, allowing to make the cpufreq support work when the CPU clock source are not the default ones.

Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
[gregory: extract from a larger patch, modify comments and commit log]
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/cpufreq/armada-37xx-cpufreq.c +17 -3
index 75491fc841a6..ad4463e4266e 100644
@@ -162,11 +162,25 @@ static void __init armada37xx_cpufreq_dvfs_setup(struct regmap *base,
 	}
 
 	/*
-	 * Set cpu clock source, for all the level we keep the same
-	 * clock source that the one already configured. For this one
-	 * we need to use the clock framework
+	 * Set CPU clock source, for all the level we keep the same
+	 * clock source that the one already configured with DVS
+	 * disabled. For this one we need to use the clock framework
 	 */
 	parent = clk_get_parent(clk);
+
+	/*
+	 * Unset parent clock to force the clock framework setting again
+	 * the clock parent
+	 */
+	clk_set_parent(clk, NULL);
+
+	/*
+	 * For the Armada 37xx CPU clocks, setting the parent will
+	 * actually configure the parent when DVFS is enabled. At
+	 * hardware level it will be a different register from the one
+	 * read when doing clk_get_parent that will be set with
+	 * clk_set_parent.
+	 */
 	clk_set_parent(clk, parent);
 }
 
Signature · 1 line
2.20.1

I applied this and selected only
CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE=y
CONFIG_CPU_FREQ_GOV_PERFORMANCE=y
CONFIG_CPU_FREQ_GOV_POWERSAVE=y

The board boots up with 'powersave' instead of 'performance' (maybe some init script?)

When i tried changing to 'performance' i got this:

[  164.134589] Internal error: synchronous parity or ECC error: 96000018 [#1]
PREEMPT SMP
[  164.139899] Modules linked in: mvneta mvmdio phylink crct10dif_ce ip_tables
x_tables ipv6
[  164.148321] CPU: 1 PID: 2485 Comm: bash Not tainted 5.0.0espressobin+ #68
[  164.155305] Hardware name: Globalscale Marvell ESPRESSOBin Board (DT)
[  164.161938] pstate: 000003c5 (nzcv DAIF -PAN -UAO)
[  164.166873] pc : do_mem_abort+0x48/0xb0
[  164.170808] lr : el1_da+0x20/0x88
[  164.174210] sp : ffff000011b5b1f0
[  164.177614] x29: ffff000011b5b1f0 x28: ffff80003a0ab800 
[  164.183080] x27: 0000000000000000 x26: 0000000000000000 
[  164.188545] x25: 0000000000000000 x24: 0000000000000025 
[  164.194012] x23: 00000000000003c5 x22: ffff000010c37318 
[  164.199477] x21: ffff000010c371f0 x20: ffff000011b5b220 
[  164.204942] x19: 000000009600004f x18: 0000000000000000 
[  164.210408] x17: 0000000000000000 x16: 0000000000000005 
[  164.215873] x15: 0000000000000000 x14: 0000000000000000 
[  164.221339] x13: ffff80003ac55168 x12: ffff80003ac55168 
[  164.226805] x11: 0000000000000026 x10: 0101010101010101 
[  164.232270] x9 : 0000000000000000 x8 : 0000000000000000 
[  164.237735] x7 : 0000000000000000 x6 : 0000000000000000 
[  164.243201] x5 : 0000000000000000 x4 : 0000000000000000 
[  164.248667] x3 : 0000000000000000 x2 : 0000000000000000 
[  164.254133] x1 : 0000000000000000 x0 : 0000000000000000 
[  164.259601] Process bash (pid: 2485, stack limit = 0x00000000eb09c765)
[  164.266318] Call trace:
[  164.268830]  do_mem_abort+0x48/0xb0
[  164.272413]  el1_da+0x20/0x88
[  164.275460]  el1_error+0x14/0xe4
[  164.278778] Code: 2a1303e1 9ba07c63 aa1603e0 8b030095 (f8636883) 
[  164.285053] ---[ end trace 78a878c96fcb4581 ]---
[  164.289795] Kernel panic - not syncing: Fatal exception
[  164.295171] SMP: stopping secondary CPUs
[  164.299211] Kernel Offset: disabled
[  164.302788] CPU features: 0x002,2000200c
[  164.306817] Memory Limit: none
[  164.309961] ---[ end Kernel panic - not syncing: Fatal exception ]---

Cheers
/Ilias

author maintainer
#5 lore raw

PATCH 2/2 cpufreq: armada-37xx: fix frequency calculation for opp

In reply to Gregory CLEMENT

The frequency calculation was based on the current(max) frequency of the CPU. However for low frequency, the value used was already the parent frequency divided by a factor of 2.

Instead of using this frequency, this fix directly get the frequency from the parent clock.

Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)
drivers/cpufreq/armada-37xx-cpufreq.c +19 -3
index ad4463e4266e..a0962463805e 100644
@@ -373,11 +373,11 @@ static int __init armada37xx_cpufreq_driver_init(void)
 	struct armada_37xx_dvfs *dvfs;
 	struct platform_device *pdev;
 	unsigned long freq;
-	unsigned int cur_frequency;
+	unsigned int cur_frequency, base_frequency;
 	struct regmap *nb_pm_base, *avs_base;
 	struct device *cpu_dev;
 	int load_lvl, ret;
-	struct clk *clk;
+	struct clk *clk, *parent;
 
 	nb_pm_base =
 		syscon_regmap_lookup_by_compatible("marvell,armada-3700-nb-pm");
@@ -413,6 +413,22 @@ static int __init armada37xx_cpufreq_driver_init(void)
 		return PTR_ERR(clk);
 	}
 
+	parent = clk_get_parent(clk);
+	if (IS_ERR(parent)) {
+		dev_err(cpu_dev, "Cannot get parent clock for CPU0\n");
+		clk_put(clk);
+		return PTR_ERR(parent);
+	}
+
+	/* Get parent CPU frequency */
+	base_frequency =  clk_get_rate(parent);
+
+	if (!base_frequency) {
+		dev_err(cpu_dev, "Failed to get parent clock rate for CPU\n");
+		clk_put(clk);
+		return -EINVAL;
+	}
+
 	/* Get nominal (current) CPU frequency */
 	cur_frequency = clk_get_rate(clk);
 	if (!cur_frequency) {
@@ -445,7 +461,7 @@ static int __init armada37xx_cpufreq_driver_init(void)
 	for (load_lvl = ARMADA_37XX_DVFS_LOAD_0; load_lvl < LOAD_LEVEL_NR;
 	     load_lvl++) {
 		unsigned long u_volt = avs_map[dvfs->avs[load_lvl]] * 1000;
-		freq = cur_frequency / dvfs->divider[load_lvl];
+		freq = base_frequency / dvfs->divider[load_lvl];
 		ret = dev_pm_opp_add(cpu_dev, freq, u_volt);
 		if (ret)
 			goto remove_opp;
Signature · 1 line
2.20.1
maintainer
#6 lore raw

Re: [PATCH 2/2] cpufreq: armada-37xx: fix frequency calculation for opp

Quoted 57 lines · Gregory CLEMENT

The frequency calculation was based on the current(max) frequency of the CPU. However for low frequency, the value used was already the parent frequency divided by a factor of 2.

Instead of using this frequency, this fix directly get the frequency from the parent clock.

Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)
drivers/cpufreq/armada-37xx-cpufreq.c +18 -2
index ad4463e4266e..a0962463805e 100644
@@ -373,11 +373,11 @@ static int __init armada37xx_cpufreq_driver_init(void)
 	struct armada_37xx_dvfs *dvfs;
 	struct platform_device *pdev;
 	unsigned long freq;
-	unsigned int cur_frequency;
+	unsigned int cur_frequency, base_frequency;
 	struct regmap *nb_pm_base, *avs_base;
 	struct device *cpu_dev;
 	int load_lvl, ret;
-	struct clk *clk;
+	struct clk *clk, *parent;
 
 	nb_pm_base =
 		syscon_regmap_lookup_by_compatible("marvell,armada-3700-nb-pm");
@@ -413,6 +413,22 @@ static int __init armada37xx_cpufreq_driver_init(void)
 		return PTR_ERR(clk);
 	}
 
+	parent = clk_get_parent(clk);
+	if (IS_ERR(parent)) {
+		dev_err(cpu_dev, "Cannot get parent clock for CPU0\n");
+		clk_put(clk);
+		return PTR_ERR(parent);
+	}
+
+	/* Get parent CPU frequency */
+	base_frequency =  clk_get_rate(parent);
+
+	if (!base_frequency) {
+		dev_err(cpu_dev, "Failed to get parent clock rate for CPU\n");
+		clk_put(clk);
+		return -EINVAL;
+	}
+
 	/* Get nominal (current) CPU frequency */
 	cur_frequency = clk_get_rate(clk);
 	if (!cur_frequency) {

drivers/cpufreq/armada-37xx-cpufreq.c static int __init armada37xx_cpufreq_driver_init(void) · line 467

Quoted 8 lines
drivers/cpufreq/armada-37xx-cpufreq.c +1 -1
@@ -445,7 +461,7 @@ static int __init armada37xx_cpufreq_driver_init(void)
 	for (load_lvl = ARMADA_37XX_DVFS_LOAD_0; load_lvl < LOAD_LEVEL_NR;
 	     load_lvl++) {
 		unsigned long u_volt = avs_map[dvfs->avs[load_lvl]] * 1000;
-		freq = cur_frequency / dvfs->divider[load_lvl];
+		freq = base_frequency / dvfs->divider[load_lvl];
 		ret = dev_pm_opp_add(cpu_dev, freq, u_volt);
 		if (ret)
 			goto remove_opp;

Great. I was wondering on why the top frequency was skipped :)

Signature · 1 line
viresh
maintainer
#7 lore raw

Re: [PATCH 2/2] cpufreq: armada-37xx: fix frequency calculation for opp

In reply to Gregory CLEMENT
Quoted 7 lines · Gregory CLEMENT

The frequency calculation was based on the current(max) frequency of the CPU. However for low frequency, the value used was already the parent frequency divided by a factor of 2.

Instead of using this frequency, this fix directly get the frequency from the parent clock.

Quoted 7 lines
Fixes 92ce45fb875d ("cpufreq: Add DVFS support for Armada 37xx")
---
 drivers/cpufreq/armada-37xx-cpufreq.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)

Applied this one.

Signature · 1 line
viresh