cpufreq: qoriq: optimize the CPU frequency switching time

Commit Message

tang yuantian June 4, 2015, 6:25 a.m. UTC

From: Tang Yuantian <[e-mail]>

Each time the CPU switches its frequency, the clock nodes in
DTS are walked through to find proper clock source. This is
very time-consuming, for example, it is up to 500+ us on T4240.
Besides, switching time varies from clock to clock.
To optimize this, each input clock of CPU is buffered, so that
it can be picked up instantly when needed.

Since for each CPU each input clock is stored in a pointer
which takes 4 or 8 bytes memory and normally there are several
input clocks per CPU, that will not take much memory as well.

Signed-off-by: Tang Yuantian <[e-mail]>
---
 drivers/cpufreq/qoriq-cpufreq.c | 32 +++++++++++++++++++++-----------
 1 file changed, 21 insertions(+), 11 deletions(-)

Comments

Viresh Kumar June 4, 2015, 6:36 a.m. UTC | #1

On 04-06-15, 14:25, [e-mail] wrote:
> From: Tang Yuantian <[e-mail]>
> 
> Each time the CPU switches its frequency, the clock nodes in
> DTS are walked through to find proper clock source. This is
> very time-consuming, for example, it is up to 500+ us on T4240.
> Besides, switching time varies from clock to clock.
> To optimize this, each input clock of CPU is buffered, so that
> it can be picked up instantly when needed.
> 
> Since for each CPU each input clock is stored in a pointer
> which takes 4 or 8 bytes memory and normally there are several
> input clocks per CPU, that will not take much memory as well.

Not sure how it got included in this form in the first place. :)

> Signed-off-by: Tang Yuantian <[e-mail]>
> ---
>  drivers/cpufreq/qoriq-cpufreq.c | 32 +++++++++++++++++++++-----------
>  1 file changed, 21 insertions(+), 11 deletions(-)

Acked-by: Viresh Kumar <[e-mail]>

Rafael J. Wysocki June 15, 2015, 11:46 p.m. UTC | #2

On Thursday, June 04, 2015 12:06:36 PM Viresh Kumar wrote:
> On 04-06-15, 14:25, [e-mail] wrote:
> > From: Tang Yuantian <[e-mail]>
> > 
> > Each time the CPU switches its frequency, the clock nodes in
> > DTS are walked through to find proper clock source. This is
> > very time-consuming, for example, it is up to 500+ us on T4240.
> > Besides, switching time varies from clock to clock.
> > To optimize this, each input clock of CPU is buffered, so that
> > it can be picked up instantly when needed.
> > 
> > Since for each CPU each input clock is stored in a pointer
> > which takes 4 or 8 bytes memory and normally there are several
> > input clocks per CPU, that will not take much memory as well.
> 
> Not sure how it got included in this form in the first place. :)
> 
> > Signed-off-by: Tang Yuantian <[e-mail]>
> > ---
> >  drivers/cpufreq/qoriq-cpufreq.c | 32 +++++++++++++++++++++-----------
> >  1 file changed, 21 insertions(+), 11 deletions(-)
> 
> Acked-by: Viresh Kumar <[e-mail]>

Queued up for 4.2, thanks!

Patch

diff --git a/drivers/cpufreq/qoriq-cpufreq.c b/drivers/cpufreq/qoriq-cpufreq.c
index 88b21ae..358f075 100644
--- a/drivers/cpufreq/qoriq-cpufreq.c
+++ b/drivers/cpufreq/qoriq-cpufreq.c
@@ -27,11 +27,11 @@ 
 
 /**
  * struct cpu_data
- * @parent: the parent node of cpu clock
+ * @pclk: the parent clock of cpu
  * @table: frequency table
  */
 struct cpu_data {
-	struct device_node *parent;
+	struct clk **pclk;
 	struct cpufreq_frequency_table *table;
 };
 
@@ -196,7 +196,7 @@  static void freq_table_sort(struct cpufreq_frequency_table *freq_table,
 
 static int qoriq_cpufreq_cpu_init(struct cpufreq_policy *policy)
 {
-	struct device_node *np;
+	struct device_node *np, *pnode;
 	int i, count, ret;
 	u32 freq, mask;
 	struct clk *clk;
@@ -219,17 +219,23 @@  static int qoriq_cpufreq_cpu_init(struct cpufreq_policy *policy)
 		goto err_nomem2;
 	}
 
-	data->parent = of_parse_phandle(np, "clocks", 0);
-	if (!data->parent) {
+	pnode = of_parse_phandle(np, "clocks", 0);
+	if (!pnode) {
 		pr_err("%s: could not get clock information\n", __func__);
 		goto err_nomem2;
 	}
 
-	count = of_property_count_strings(data->parent, "clock-names");
+	count = of_property_count_strings(pnode, "clock-names");
+	data->pclk = kcalloc(count, sizeof(struct clk *), GFP_KERNEL);
+	if (!data->pclk) {
+		pr_err("%s: no memory\n", __func__);
+		goto err_node;
+	}
+
 	table = kcalloc(count + 1, sizeof(*table), GFP_KERNEL);
 	if (!table) {
 		pr_err("%s: no memory\n", __func__);
-		goto err_node;
+		goto err_pclk;
 	}
 
 	if (fmask)
@@ -238,7 +244,8 @@  static int qoriq_cpufreq_cpu_init(struct cpufreq_policy *policy)
 		mask = 0x0;
 
 	for (i = 0; i < count; i++) {
-		clk = of_clk_get(data->parent, i);
+		clk = of_clk_get(pnode, i);
+		data->pclk[i] = clk;
 		freq = clk_get_rate(clk);
 		/*
 		 * the clock is valid if its frequency is not masked
@@ -273,13 +280,16 @@  static int qoriq_cpufreq_cpu_init(struct cpufreq_policy *policy)
 	policy->cpuinfo.transition_latency = u64temp + 1;
 
 	of_node_put(np);
+	of_node_put(pnode);
 
 	return 0;
 
 err_nomem1:
 	kfree(table);
+err_pclk:
+	kfree(data->pclk);
 err_node:
-	of_node_put(data->parent);
+	of_node_put(pnode);
 err_nomem2:
 	policy->driver_data = NULL;
 	kfree(data);
@@ -293,7 +303,7 @@  static int __exit qoriq_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 {
 	struct cpu_data *data = policy->driver_data;
 
-	of_node_put(data->parent);
+	kfree(data->pclk);
 	kfree(data->table);
 	kfree(data);
 	policy->driver_data = NULL;
@@ -307,7 +317,7 @@  static int qoriq_cpufreq_target(struct cpufreq_policy *policy,
 	struct clk *parent;
 	struct cpu_data *data = policy->driver_data;
 
-	parent = of_clk_get(data->parent, data->table[index].driver_data);
+	parent = data->pclk[data->table[index].driver_data];
 	return clk_set_parent(policy->clk, parent);
 }