acpi-cpufreq: replace per_cpu with driver_data of policy

Drivers can store their internal per-policy information in policy->driver_data, lets use it.

we have benefits after this replacing.
1) memory saving.
2) policy is shared by several cpus, per_cpu seems not correct. using *driver_data* is more reasonable.
3) fix a memory leak in acpi_cpufreq_cpu_exit. as policy->cpu might change during cpu hotplug. So sometimes we cant't free *data*, use *driver_data* to fix it.
4) fix a zero return value of get_cur_freq_on_cpu. Only per_cpu of policy->cpu is set to *data*, if we try to get cpufreq on other cpus, we get zero instead of correct values. Use *driver_data* to fix it.

Signed-off-by Pan Xinhui <[email protected]>

---
Changes from V1:
codes style fix, comments update
move cpufreq_cpu_put(policy) after we get *driver_data*

---
 drivers/cpufreq/acpi-cpufreq.c | 40 ++++++++++++++++++++++------------------
 1 file changed, 22 insertions(+), 18 deletions(-)

drivers/cpufreq/acpi-cpufreq.c +22 -18

index 0136dfc..e7fcaa6 100644
@@ -72,8 +72,6 @@ struct acpi_cpufreq_data {
 	cpumask_var_t freqdomain_cpus;
 };
 
-static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
-
 /* acpi_perf_data is a pointer to percpu data. */
 static struct acpi_processor_performance __percpu *acpi_perf_data;
 

@@ -144,7 +142,7 @@ static int _store_boost(int val)
 
 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
+	struct acpi_cpufreq_data *data = policy->driver_data;
 
 	return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 }

@@ -327,7 +325,8 @@ static void drv_write(struct drv_cmd *cmd)
 	put_cpu();
 }
 
-static u32 get_cur_val(const struct cpumask *mask)
+static u32
+get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 {
 	struct acpi_processor_performance *perf;
 	struct drv_cmd cmd;

@@ -335,7 +334,7 @@ static u32 get_cur_val(const struct cpumask *mask)
 	if (unlikely(cpumask_empty(mask)))
 		return 0;
 
-	switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
+	switch (data->cpu_feature) {
 	case SYSTEM_INTEL_MSR_CAPABLE:
 		cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 		cmd.addr.msr.reg = MSR_IA32_PERF_CTL;

@@ -346,7 +345,7 @@ static u32 get_cur_val(const struct cpumask *mask)
 		break;
 	case SYSTEM_IO_CAPABLE:
 		cmd.type = SYSTEM_IO_CAPABLE;
-		perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
+		perf = data->acpi_data;
 		cmd.addr.io.port = perf->control_register.address;
 		cmd.addr.io.bit_width = perf->control_register.bit_width;
 		break;

@@ -364,19 +363,24 @@ static u32 get_cur_val(const struct cpumask *mask)
 
 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
+	struct acpi_cpufreq_data *data;
+	struct cpufreq_policy *policy;
 	unsigned int freq;
 	unsigned int cached_freq;
 
 	pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 
-	if (unlikely(data == NULL ||
-		     data->acpi_data == NULL || data->freq_table == NULL)) {
+	policy = cpufreq_cpu_get(cpu);
+	if (unlikely(!policy))
+		return 0;
+
+	data = policy->driver_data;
+	cpufreq_cpu_put(policy);
+	if (unlikely(!data || !data->acpi_data || !data->freq_table))
 		return 0;
-	}
 
 	cached_freq = data->freq_table[data->acpi_data->state].frequency;
-	freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
+	freq = extract_freq(get_cur_val(cpumask_of(cpu), data), data);
 	if (freq != cached_freq) {
 		/*
 		 * The dreaded BIOS frequency change behind our back.

@@ -397,7 +401,7 @@ static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 	unsigned int i;
 
 	for (i = 0; i < 100; i++) {
-		cur_freq = extract_freq(get_cur_val(mask), data);
+		cur_freq = extract_freq(get_cur_val(mask, data), data);
 		if (cur_freq == freq)
 			return 1;
 		udelay(10);

@@ -408,7 +412,7 @@ static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 			       unsigned int index)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
+	struct acpi_cpufreq_data *data = policy->driver_data;
 	struct acpi_processor_performance *perf;
 	struct drv_cmd cmd;
 	unsigned int next_perf_state = 0; /* Index into perf table */

@@ -673,7 +677,7 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 	}
 
 	data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
-	per_cpu(acfreq_data, cpu) = data;
+	policy->driver_data = data;
 
 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;

@@ -843,19 +847,19 @@ err_free_mask:
 	free_cpumask_var(data->freqdomain_cpus);
 err_free:
 	kfree(data);
-	per_cpu(acfreq_data, cpu) = NULL;
+	policy->driver_data = NULL;
 
 	return result;
 }
 
 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
+	struct acpi_cpufreq_data *data = policy->driver_data;
 
 	pr_debug("acpi_cpufreq_cpu_exit\n");
 
 	if (data) {
-		per_cpu(acfreq_data, policy->cpu) = NULL;
+		policy->driver_data = NULL;
 		acpi_processor_unregister_performance(data->acpi_data,
 						      policy->cpu);
 		free_cpumask_var(data->freqdomain_cpus);

@@ -868,7 +872,7 @@ static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 
 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
+	struct acpi_cpufreq_data *data = policy->driver_data;
 
 	pr_debug("acpi_cpufreq_resume\n");
 

Drivers can store their internal per-policy information in policy->driver_data, lets use it.

we have benefits after this replacing.
1) memory saving.
2) policy is shared by several cpus, per_cpu seems not correct. using *driver_data* is more reasonable.
3) fix a memory leak in acpi_cpufreq_cpu_exit. as policy->cpu might change during cpu hotplug. So sometimes we cant't free *data*, use *driver_data* to fix it.
4) fix a zero return value of get_cur_freq_on_cpu. Only per_cpu of policy->cpu is set to *data*, if we try to get cpufreq on other cpus, we get zero instead of correct values. Use *driver_data* to fix it.

Signed-off-by Pan Xinhui <[email protected]>

---
Changes from V1:
codes style fix, comments update
move cpufreq_cpu_put(policy) after we get *driver_data*

---
 drivers/cpufreq/acpi-cpufreq.c | 40 ++++++++++++++++++++++------------------
 1 file changed, 22 insertions(+), 18 deletions(-)

Quoted 15 lines · Pan Xinhui

Drivers can store their internal per-policy information in policy->driver_data, lets use it.

we have benefits after this replacing.
1) memory saving.
2) policy is shared by several cpus, per_cpu seems not correct. using *driver_data* is more reasonable.
3) fix a memory leak in acpi_cpufreq_cpu_exit. as policy->cpu might change during cpu hotplug. So sometimes we cant't free *data*, use *driver_data* to fix it.
4) fix a zero return value of get_cur_freq_on_cpu. Only per_cpu of policy->cpu is set to *data*, if we try to get cpufreq on other cpus, we get zero instead of correct values. Use *driver_data* to fix it.

Quoted 8 lines

Signed-off-by Pan Xinhui <[email protected]>

---
Changes from V1:
codes style fix, comments update
move cpufreq_cpu_put(policy) after we get *driver_data*

---
 drivers/cpufreq/acpi-cpufreq.c | 40 ++++++++++++++++++++++------------------
 1 file changed, 22 insertions(+), 18 deletions(-)

Acked-by Viresh Kumar <[email protected]>

Quoted 25 lines · Viresh Kumar

Quoted 23 lines · Pan Xinhui

Drivers can store their internal per-policy information in policy->driver_data, lets use it.

we have benefits after this replacing.
1) memory saving.
2) policy is shared by several cpus, per_cpu seems not correct. using *driver_data* is more reasonable.
3) fix a memory leak in acpi_cpufreq_cpu_exit. as policy->cpu might change during cpu hotplug. So sometimes we cant't free *data*, use *driver_data* to fix it.
4) fix a zero return value of get_cur_freq_on_cpu. Only per_cpu of policy->cpu is set to *data*, if we try to get cpufreq on other cpus, we get zero instead of correct values. Use *driver_data* to fix it.

Signed-off-by Pan Xinhui <[email protected]>

---
Changes from V1:
codes style fix, comments update
move cpufreq_cpu_put(policy) after we get *driver_data*

---
 drivers/cpufreq/acpi-cpufreq.c | 40 ++++++++++++++++++++++------------------
 1 file changed, 22 insertions(+), 18 deletions(-)

Quoted 1 line

Acked-by Viresh Kumar <[email protected]>

OK

Does it fix any recent regressions or is it just an old bug?

Rafael

hi, Rafael
thanks for your kind reply. :)

Quoted 30 lines · Rafael J. Wysocki

Quoted 25 lines · Viresh Kumar

Quoted 15 lines · Pan Xinhui

Drivers can store their internal per-policy information in policy->driver_data, lets use it.

we have benefits after this replacing.
1) memory saving.
2) policy is shared by several cpus, per_cpu seems not correct. using *driver_data* is more reasonable.
3) fix a memory leak in acpi_cpufreq_cpu_exit. as policy->cpu might change during cpu hotplug. So sometimes we cant't free *data*, use *driver_data* to fix it.
4) fix a zero return value of get_cur_freq_on_cpu. Only per_cpu of policy->cpu is set to *data*, if we try to get cpufreq on other cpus, we get zero instead of correct values. Use *driver_data* to fix it.

Quoted 8 lines

Signed-off-by Pan Xinhui <[email protected]>

---
Changes from V1:
codes style fix, comments update
move cpufreq_cpu_put(policy) after we get *driver_data*

---
 drivers/cpufreq/acpi-cpufreq.c | 40 ++++++++++++++++++++++------------------
 1 file changed, 22 insertions(+), 18 deletions(-)

Quoted 1 line

Acked-by Viresh Kumar <[email protected]>

OK

Quoted 2 lines

Does it fix any recent regressions or is it just an old bug?

This patch achieve old bug fix and codes improvements. In past days, policy has no field *driver_data*, So acpi-cpufreq driver has to use per_cpu to store some extra information. But it did not take good care of every scenarios. Now cpufreq core makes awesome effort to store more per-policy information in policy. We can make use of this feature. So I cook this patch. :)

I am preparing two patches for other two issues in acpi-cpufreq driver based on this patch. I will fix them step by step. :)

drivers/cpufreq/acpi-cpufreq.c +2 -1

@@ -364,19 +363,24 @@ static u32 get_cur_val(const struct cpumask *mask)
 
 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
+	struct acpi_cpufreq_data *data;
+	struct cpufreq_policy *policy;
 	unsigned int freq;
 	unsigned int cached_freq;
 

drivers/cpufreq/acpi-cpufreq.c static u32 get_cur_val(const struct cpumask *mask) · line 378

Quoted 2 lines

drivers/cpufreq/acpi-cpufreq.c +2 -0

+	data = policy->driver_data;
+	cpufreq_cpu_put(policy);

If we put policy here can we guarantee that memory pointed to by data stays valid? Shoudln't we issue cpufreq_cpu_put(policy) after we done assessing the pointer?

Hi Pan,

Quoted 12 lines · Pan Xinhui

drivers/cpufreq/acpi-cpufreq.c +2 -1

@@ -364,19 +363,24 @@ static u32 get_cur_val(const struct cpumask *mask)
 
 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 {
-	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
+	struct acpi_cpufreq_data *data;
+	struct cpufreq_policy *policy;
 	unsigned int freq;
 	unsigned int cached_freq;
 
 	pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 

drivers/cpufreq/acpi-cpufreq.c static u32 get_cur_val(const struct cpumask *mask) · line 378

Quoted 8 lines

drivers/cpufreq/acpi-cpufreq.c +6 -2

-	if (unlikely(data == NULL ||
-		     data->acpi_data == NULL || data->freq_table == NULL)) {
+	policy = cpufreq_cpu_get(cpu);
+	if (unlikely(!policy))
+		return 0;
+
+	data = policy->driver_data;
+	cpufreq_cpu_put(policy);

If we put policy here can we guarantee that memory pointed to by data stays valid? Shoudln't we issue cpufreq_cpu_put(policy) after we done assessing the pointer?

drivers/cpufreq/acpi-cpufreq.c static u32 get_cur_val(const struct cpumask *mask) · line 386

Quoted 10 lines

drivers/cpufreq/acpi-cpufreq.c +2 -2

+	if (unlikely(!data || !data->acpi_data || !data->freq_table))
 		return 0;
-	}
 
 	cached_freq = data->freq_table[data->acpi_data->state].frequency;
-	freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
+	freq = extract_freq(get_cur_val(cpumask_of(cpu), data), data);
 	if (freq != cached_freq) {
 		/*
 		 * The dreaded BIOS frequency change behind our back.

Thanks.