P-state software coordination for speedstep-centrino

http://bugzilla.kernel.org/show_bug.cgi?id=5737

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>

1 files changed, 180 insertions, 66 deletions

diff --git a/arch/i386/kernel/cpu/cpufreq/speedstep-centrino.c b/arch/i386/kernel/cpu/cpufreq/speedstep-centrino.c
index c173c0fa117a..37dee862f0d3 100644
--- a/arch/i386/kernel/cpu/cpufreq/speedstep-centrino.c
+++ b/arch/i386/kernel/cpu/cpufreq/speedstep-centrino.c
@@ -351,7 +351,36 @@ static unsigned int get_cur_freq(unsigned int cpu)
 
 #ifdef CONFIG_X86_SPEEDSTEP_CENTRINO_ACPI
 
-static struct acpi_processor_performance p;
+static struct acpi_processor_performance *acpi_perf_data[NR_CPUS];
+
+/*
+ * centrino_cpu_early_init_acpi - Do the preregistering with ACPI P-States
+ * library
+ *
+ * Before doing the actual init, we need to do _PSD related setup whenever
+ * supported by the BIOS. These are handled by this early_init routine.
+ */
+static int centrino_cpu_early_init_acpi(void)
+{
+        unsigned int    i, j;
+        struct acpi_processor_performance       *data;
+
+        for_each_cpu(i) {
+                data = kzalloc(sizeof(struct acpi_processor_performance), 
+                                GFP_KERNEL);
+                if (!data) {
+                        for_each_cpu(j) {
+                                kfree(acpi_perf_data[j]);
+                                acpi_perf_data[j] = NULL;
+                        }
+                        return (-ENOMEM);
+                }
+                acpi_perf_data[i] = data;
+        }
+
+        acpi_processor_preregister_performance(acpi_perf_data);
+        return 0;
+}
 
 /*
  * centrino_cpu_init_acpi - register with ACPI P-States library
@@ -365,46 +394,51 @@ static int centrino_cpu_init_acpi(struct cpufreq_policy *policy)
         unsigned long                   cur_freq;
         int                             result = 0, i;
         unsigned int                    cpu = policy->cpu;
+        struct acpi_processor_performance       *p;
+
+        p = acpi_perf_data[cpu];
 
         /* register with ACPI core */
-        if (acpi_processor_register_performance(&p, cpu)) {
+        if (acpi_processor_register_performance(p, cpu)) {
                 dprintk(KERN_INFO PFX "obtaining ACPI data failed\n");
                 return -EIO;
         }
+        policy->cpus = p->shared_cpu_map;
+        policy->shared_type = p->shared_type;
 
         /* verify the acpi_data */
-        if (p.state_count <= 1) {
+        if (p->state_count <= 1) {
                 dprintk("No P-States\n");
                 result = -ENODEV;
                 goto err_unreg;
         }
 
-        if ((p.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
-            (p.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
+        if ((p->control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
+            (p->status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
                 dprintk("Invalid control/status registers (%x - %x)\n",
-                        p.control_register.space_id, p.status_register.space_id);
+                        p->control_register.space_id, p->status_register.space_id);
                 result = -EIO;
                 goto err_unreg;
         }
 
-        for (i=0; i<p.state_count; i++) {
-                if (p.states[i].control != p.states[i].status) {
+        for (i=0; i<p->state_count; i++) {
+                if (p->states[i].control != p->states[i].status) {
                         dprintk("Different control (%llu) and status values (%llu)\n",
-                                p.states[i].control, p.states[i].status);
+                                p->states[i].control, p->states[i].status);
                         result = -EINVAL;
                         goto err_unreg;
                 }
 
-                if (!p.states[i].core_frequency) {
+                if (!p->states[i].core_frequency) {
                         dprintk("Zero core frequency for state %u\n", i);
                         result = -EINVAL;
                         goto err_unreg;
                 }
 
-                if (p.states[i].core_frequency > p.states[0].core_frequency) {
+                if (p->states[i].core_frequency > p->states[0].core_frequency) {
                         dprintk("P%u has larger frequency (%llu) than P0 (%llu), skipping\n", i,
-                                p.states[i].core_frequency, p.states[0].core_frequency);
-                        p.states[i].core_frequency = 0;
+                                p->states[i].core_frequency, p->states[0].core_frequency);
+                        p->states[i].core_frequency = 0;
                         continue;
                 }
         }
@@ -416,26 +450,26 @@ static int centrino_cpu_init_acpi(struct cpufreq_policy *policy)
         }
 
         centrino_model[cpu]->model_name=NULL;
-        centrino_model[cpu]->max_freq = p.states[0].core_frequency * 1000;
+        centrino_model[cpu]->max_freq = p->states[0].core_frequency * 1000;
         centrino_model[cpu]->op_points =  kmalloc(sizeof(struct cpufreq_frequency_table) *
-                                             (p.state_count + 1), GFP_KERNEL);
+                                             (p->state_count + 1), GFP_KERNEL);
         if (!centrino_model[cpu]->op_points) {
                 result = -ENOMEM;
                 goto err_kfree;
         }
 
-        for (i=0; i<p.state_count; i++) {
-                centrino_model[cpu]->op_points[i].index = p.states[i].control;
-                centrino_model[cpu]->op_points[i].frequency = p.states[i].core_frequency * 1000;
+        for (i=0; i<p->state_count; i++) {
+                centrino_model[cpu]->op_points[i].index = p->states[i].control;
+                centrino_model[cpu]->op_points[i].frequency = p->states[i].core_frequency * 1000;
                 dprintk("adding state %i with frequency %u and control value %04x\n", 
                         i, centrino_model[cpu]->op_points[i].frequency, centrino_model[cpu]->op_points[i].index);
         }
-        centrino_model[cpu]->op_points[p.state_count].frequency = CPUFREQ_TABLE_END;
+        centrino_model[cpu]->op_points[p->state_count].frequency = CPUFREQ_TABLE_END;
 
         cur_freq = get_cur_freq(cpu);
 
-        for (i=0; i<p.state_count; i++) {
-                if (!p.states[i].core_frequency) {
+        for (i=0; i<p->state_count; i++) {
+                if (!p->states[i].core_frequency) {
                         dprintk("skipping state %u\n", i);
                         centrino_model[cpu]->op_points[i].frequency = CPUFREQ_ENTRY_INVALID;
                         continue;
@@ -451,7 +485,7 @@ static int centrino_cpu_init_acpi(struct cpufreq_policy *policy)
                 }
 
                 if (cur_freq == centrino_model[cpu]->op_points[i].frequency)
-                        p.state = i;
+                        p->state = i;
         }
 
         /* notify BIOS that we exist */
@@ -464,12 +498,13 @@ static int centrino_cpu_init_acpi(struct cpufreq_policy *policy)
  err_kfree:
         kfree(centrino_model[cpu]);
  err_unreg:
-        acpi_processor_unregister_performance(&p, cpu);
+        acpi_processor_unregister_performance(p, cpu);
         dprintk(KERN_INFO PFX "invalid ACPI data\n");
         return (result);
 }
 #else
 static inline int centrino_cpu_init_acpi(struct cpufreq_policy *policy) { return -ENODEV; }
+static inline int centrino_cpu_early_init_acpi(void) { return 0; }
 #endif
 
 static int centrino_cpu_init(struct cpufreq_policy *policy)
@@ -557,10 +592,15 @@ static int centrino_cpu_exit(struct cpufreq_policy *policy)
 
 #ifdef CONFIG_X86_SPEEDSTEP_CENTRINO_ACPI
         if (!centrino_model[cpu]->model_name) {
-                dprintk("unregistering and freeing ACPI data\n");
-                acpi_processor_unregister_performance(&p, cpu);
-                kfree(centrino_model[cpu]->op_points);
-                kfree(centrino_model[cpu]);
+                static struct acpi_processor_performance *p;
+
+                if (acpi_perf_data[cpu]) {
+                        p = acpi_perf_data[cpu];
+                        dprintk("unregistering and freeing ACPI data\n");
+                        acpi_processor_unregister_performance(p, cpu);
+                        kfree(centrino_model[cpu]->op_points);
+                        kfree(centrino_model[cpu]);
+                }
         }
 #endif
 
@@ -594,63 +634,124 @@ static int centrino_target (struct cpufreq_policy *policy,
                             unsigned int relation)
 {
         unsigned int    newstate = 0;
-        unsigned int    msr, oldmsr, h, cpu = policy->cpu;
+        unsigned int    msr, oldmsr = 0, h = 0, cpu = policy->cpu;
         struct cpufreq_freqs    freqs;
+        cpumask_t               online_policy_cpus;
         cpumask_t               saved_mask;
-        int                     retval;
+        cpumask_t               set_mask;
+        cpumask_t               covered_cpus;
+        int                     retval = 0;
+        unsigned int            j, k, first_cpu, tmp;
 
-        if (centrino_model[cpu] == NULL)
+        if (unlikely(centrino_model[cpu] == NULL))
                 return -ENODEV;
 
-        /*
-         * Support for SMP systems.
-         * Make sure we are running on the CPU that wants to change frequency
-         */
-        saved_mask = current->cpus_allowed;
-        set_cpus_allowed(current, policy->cpus);
-        if (!cpu_isset(smp_processor_id(), policy->cpus)) {
-                dprintk("couldn't limit to CPUs in this domain\n");
-                return(-EAGAIN);
+        if (unlikely(cpufreq_frequency_table_target(policy,
+                        centrino_model[cpu]->op_points,
+                        target_freq,
+                        relation,
+                        &newstate))) {
+                return -EINVAL;
         }
 
-        if (cpufreq_frequency_table_target(policy, centrino_model[cpu]->op_points, target_freq,
-                                           relation, &newstate)) {
-                retval = -EINVAL;
-                goto migrate_end;
-        }
+        /* cpufreq holds the hotplug lock, so we are safe from here on */
+        cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
 
-        msr = centrino_model[cpu]->op_points[newstate].index;
-        rdmsr(MSR_IA32_PERF_CTL, oldmsr, h);
+        saved_mask = current->cpus_allowed;
+        first_cpu = 1;
+        cpus_clear(covered_cpus);
+        for_each_cpu_mask(j, online_policy_cpus) {
+                /*
+                 * Support for SMP systems.
+                 * Make sure we are running on CPU that wants to change freq
+                 */
+                cpus_clear(set_mask);
+                if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY)
+                        cpus_or(set_mask, set_mask, online_policy_cpus);
+                else
+                        cpu_set(j, set_mask);
+
+                set_cpus_allowed(current, set_mask);
+                if (unlikely(!cpu_isset(smp_processor_id(), set_mask))) {
+                        dprintk("couldn't limit to CPUs in this domain\n");
+                        retval = -EAGAIN;
+                        if (first_cpu) {
+                                /* We haven't started the transition yet. */
+                                goto migrate_end;
+                        }
+                        break;
+                }
 
-        if (msr == (oldmsr & 0xffff)) {
-                retval = 0;
-                dprintk("no change needed - msr was and needs to be %x\n", oldmsr);
-                goto migrate_end;
-        }
+                msr = centrino_model[cpu]->op_points[newstate].index;
+
+                if (first_cpu) {
+                        rdmsr(MSR_IA32_PERF_CTL, oldmsr, h);
+                        if (msr == (oldmsr & 0xffff)) {
+                                dprintk("no change needed - msr was and needs "
+                                        "to be %x\n", oldmsr);
+                                retval = 0;
+                                goto migrate_end;
+                        }
+
+                        freqs.old = extract_clock(oldmsr, cpu, 0);
+                        freqs.new = extract_clock(msr, cpu, 0);
+
+                        dprintk("target=%dkHz old=%d new=%d msr=%04x\n",
+                                target_freq, freqs.old, freqs.new, msr);
+
+                        for_each_cpu_mask(k, online_policy_cpus) {
+                                freqs.cpu = k;
+                                cpufreq_notify_transition(&freqs,
+                                        CPUFREQ_PRECHANGE);
+                        }
+
+                        first_cpu = 0;
+                        /* all but 16 LSB are reserved, treat them with care */
+                        oldmsr &= ~0xffff;
+                        msr &= 0xffff;
+                        oldmsr |= msr;
+                }
 
-        freqs.cpu = cpu;
-        freqs.old = extract_clock(oldmsr, cpu, 0);
-        freqs.new = extract_clock(msr, cpu, 0);
+                wrmsr(MSR_IA32_PERF_CTL, oldmsr, h);
+                if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY)
+                        break;
 
-        dprintk("target=%dkHz old=%d new=%d msr=%04x\n",
-                target_freq, freqs.old, freqs.new, msr);
+                cpu_set(j, covered_cpus);
+        }
 
-        cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+        for_each_cpu_mask(k, online_policy_cpus) {
+                freqs.cpu = k;
+                cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+        }
 
-        /* all but 16 LSB are "reserved", so treat them with
-           care */
-        oldmsr &= ~0xffff;
-        msr &= 0xffff;
-        oldmsr |= msr;
+        if (unlikely(retval)) {
+                /*
+                 * We have failed halfway through the frequency change.
+                 * We have sent callbacks to policy->cpus and
+                 * MSRs have already been written on coverd_cpus.
+                 * Best effort undo..
+                 */
 
-        wrmsr(MSR_IA32_PERF_CTL, oldmsr, h);
+                if (!cpus_empty(covered_cpus)) {
+                        for_each_cpu_mask(j, covered_cpus) {
+                                set_cpus_allowed(current, cpumask_of_cpu(j));
+                                wrmsr(MSR_IA32_PERF_CTL, oldmsr, h);
+                        }
+                }
 
-        cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+                tmp = freqs.new;
+                freqs.new = freqs.old;
+                freqs.old = tmp;
+                for_each_cpu_mask(j, online_policy_cpus) {
+                        freqs.cpu = j;
+                        cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+                        cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+                }
+        }
 
-        retval = 0;
 migrate_end:
         set_cpus_allowed(current, saved_mask);
-        return (retval);
+        return 0;
 }
 
 static struct freq_attr* centrino_attr[] = {
@@ -692,12 +793,25 @@ static int __init centrino_init(void)
         if (!cpu_has(cpu, X86_FEATURE_EST))
                 return -ENODEV;
 
+        centrino_cpu_early_init_acpi();
+
         return cpufreq_register_driver(&centrino_driver);
 }
 
 static void __exit centrino_exit(void)
 {
+#ifdef CONFIG_X86_SPEEDSTEP_CENTRINO_ACPI
+        unsigned int j;
+#endif
+        
         cpufreq_unregister_driver(&centrino_driver);
+
+#ifdef CONFIG_X86_SPEEDSTEP_CENTRINO_ACPI
+        for_each_cpu(j) {
+                kfree(acpi_perf_data[j]);
+                acpi_perf_data[j] = NULL;
+        }
+#endif
 }
 
 MODULE_AUTHOR ("Jeremy Fitzhardinge <jeremy@goop.org>");