Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / oprofile / oprofile_perf.c
diff --git a/kernel/drivers/oprofile/oprofile_perf.c b/kernel/drivers/oprofile/oprofile_perf.c
new file mode 100644 (file)
index 0000000..d5b2732
--- /dev/null
@@ -0,0 +1,327 @@
+/*
+ * Copyright 2010 ARM Ltd.
+ * Copyright 2012 Advanced Micro Devices, Inc., Robert Richter
+ *
+ * Perf-events backend for OProfile.
+ */
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+#include <linux/oprofile.h>
+#include <linux/slab.h>
+
+/*
+ * Per performance monitor configuration as set via oprofilefs.
+ */
+struct op_counter_config {
+       unsigned long count;
+       unsigned long enabled;
+       unsigned long event;
+       unsigned long unit_mask;
+       unsigned long kernel;
+       unsigned long user;
+       struct perf_event_attr attr;
+};
+
+static int oprofile_perf_enabled;
+static DEFINE_MUTEX(oprofile_perf_mutex);
+
+static struct op_counter_config *counter_config;
+static DEFINE_PER_CPU(struct perf_event **, perf_events);
+static int num_counters;
+
+/*
+ * Overflow callback for oprofile.
+ */
+static void op_overflow_handler(struct perf_event *event,
+                       struct perf_sample_data *data, struct pt_regs *regs)
+{
+       int id;
+       u32 cpu = smp_processor_id();
+
+       for (id = 0; id < num_counters; ++id)
+               if (per_cpu(perf_events, cpu)[id] == event)
+                       break;
+
+       if (id != num_counters)
+               oprofile_add_sample(regs, id);
+       else
+               pr_warning("oprofile: ignoring spurious overflow "
+                               "on cpu %u\n", cpu);
+}
+
+/*
+ * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile
+ * settings in counter_config. Attributes are created as `pinned' events and
+ * so are permanently scheduled on the PMU.
+ */
+static void op_perf_setup(void)
+{
+       int i;
+       u32 size = sizeof(struct perf_event_attr);
+       struct perf_event_attr *attr;
+
+       for (i = 0; i < num_counters; ++i) {
+               attr = &counter_config[i].attr;
+               memset(attr, 0, size);
+               attr->type              = PERF_TYPE_RAW;
+               attr->size              = size;
+               attr->config            = counter_config[i].event;
+               attr->sample_period     = counter_config[i].count;
+               attr->pinned            = 1;
+       }
+}
+
+static int op_create_counter(int cpu, int event)
+{
+       struct perf_event *pevent;
+
+       if (!counter_config[event].enabled || per_cpu(perf_events, cpu)[event])
+               return 0;
+
+       pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
+                                                 cpu, NULL,
+                                                 op_overflow_handler, NULL);
+
+       if (IS_ERR(pevent))
+               return PTR_ERR(pevent);
+
+       if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
+               perf_event_release_kernel(pevent);
+               pr_warning("oprofile: failed to enable event %d "
+                               "on CPU %d\n", event, cpu);
+               return -EBUSY;
+       }
+
+       per_cpu(perf_events, cpu)[event] = pevent;
+
+       return 0;
+}
+
+static void op_destroy_counter(int cpu, int event)
+{
+       struct perf_event *pevent = per_cpu(perf_events, cpu)[event];
+
+       if (pevent) {
+               perf_event_release_kernel(pevent);
+               per_cpu(perf_events, cpu)[event] = NULL;
+       }
+}
+
+/*
+ * Called by oprofile_perf_start to create active perf events based on the
+ * perviously configured attributes.
+ */
+static int op_perf_start(void)
+{
+       int cpu, event, ret = 0;
+
+       for_each_online_cpu(cpu) {
+               for (event = 0; event < num_counters; ++event) {
+                       ret = op_create_counter(cpu, event);
+                       if (ret)
+                               return ret;
+               }
+       }
+
+       return ret;
+}
+
+/*
+ * Called by oprofile_perf_stop at the end of a profiling run.
+ */
+static void op_perf_stop(void)
+{
+       int cpu, event;
+
+       for_each_online_cpu(cpu)
+               for (event = 0; event < num_counters; ++event)
+                       op_destroy_counter(cpu, event);
+}
+
+static int oprofile_perf_create_files(struct dentry *root)
+{
+       unsigned int i;
+
+       for (i = 0; i < num_counters; i++) {
+               struct dentry *dir;
+               char buf[4];
+
+               snprintf(buf, sizeof buf, "%d", i);
+               dir = oprofilefs_mkdir(root, buf);
+               oprofilefs_create_ulong(dir, "enabled", &counter_config[i].enabled);
+               oprofilefs_create_ulong(dir, "event", &counter_config[i].event);
+               oprofilefs_create_ulong(dir, "count", &counter_config[i].count);
+               oprofilefs_create_ulong(dir, "unit_mask", &counter_config[i].unit_mask);
+               oprofilefs_create_ulong(dir, "kernel", &counter_config[i].kernel);
+               oprofilefs_create_ulong(dir, "user", &counter_config[i].user);
+       }
+
+       return 0;
+}
+
+static int oprofile_perf_setup(void)
+{
+       raw_spin_lock(&oprofilefs_lock);
+       op_perf_setup();
+       raw_spin_unlock(&oprofilefs_lock);
+       return 0;
+}
+
+static int oprofile_perf_start(void)
+{
+       int ret = -EBUSY;
+
+       mutex_lock(&oprofile_perf_mutex);
+       if (!oprofile_perf_enabled) {
+               ret = 0;
+               op_perf_start();
+               oprofile_perf_enabled = 1;
+       }
+       mutex_unlock(&oprofile_perf_mutex);
+       return ret;
+}
+
+static void oprofile_perf_stop(void)
+{
+       mutex_lock(&oprofile_perf_mutex);
+       if (oprofile_perf_enabled)
+               op_perf_stop();
+       oprofile_perf_enabled = 0;
+       mutex_unlock(&oprofile_perf_mutex);
+}
+
+#ifdef CONFIG_PM
+
+static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state)
+{
+       mutex_lock(&oprofile_perf_mutex);
+       if (oprofile_perf_enabled)
+               op_perf_stop();
+       mutex_unlock(&oprofile_perf_mutex);
+       return 0;
+}
+
+static int oprofile_perf_resume(struct platform_device *dev)
+{
+       mutex_lock(&oprofile_perf_mutex);
+       if (oprofile_perf_enabled && op_perf_start())
+               oprofile_perf_enabled = 0;
+       mutex_unlock(&oprofile_perf_mutex);
+       return 0;
+}
+
+static struct platform_driver oprofile_driver = {
+       .driver         = {
+               .name           = "oprofile-perf",
+       },
+       .resume         = oprofile_perf_resume,
+       .suspend        = oprofile_perf_suspend,
+};
+
+static struct platform_device *oprofile_pdev;
+
+static int __init init_driverfs(void)
+{
+       int ret;
+
+       ret = platform_driver_register(&oprofile_driver);
+       if (ret)
+               return ret;
+
+       oprofile_pdev = platform_device_register_simple(
+                               oprofile_driver.driver.name, 0, NULL, 0);
+       if (IS_ERR(oprofile_pdev)) {
+               ret = PTR_ERR(oprofile_pdev);
+               platform_driver_unregister(&oprofile_driver);
+       }
+
+       return ret;
+}
+
+static void exit_driverfs(void)
+{
+       platform_device_unregister(oprofile_pdev);
+       platform_driver_unregister(&oprofile_driver);
+}
+
+#else
+
+static inline int  init_driverfs(void) { return 0; }
+static inline void exit_driverfs(void) { }
+
+#endif /* CONFIG_PM */
+
+void oprofile_perf_exit(void)
+{
+       int cpu, id;
+       struct perf_event *event;
+
+       for_each_possible_cpu(cpu) {
+               for (id = 0; id < num_counters; ++id) {
+                       event = per_cpu(perf_events, cpu)[id];
+                       if (event)
+                               perf_event_release_kernel(event);
+               }
+
+               kfree(per_cpu(perf_events, cpu));
+       }
+
+       kfree(counter_config);
+       exit_driverfs();
+}
+
+int __init oprofile_perf_init(struct oprofile_operations *ops)
+{
+       int cpu, ret = 0;
+
+       ret = init_driverfs();
+       if (ret)
+               return ret;
+
+       num_counters = perf_num_counters();
+       if (num_counters <= 0) {
+               pr_info("oprofile: no performance counters\n");
+               ret = -ENODEV;
+               goto out;
+       }
+
+       counter_config = kcalloc(num_counters,
+                       sizeof(struct op_counter_config), GFP_KERNEL);
+
+       if (!counter_config) {
+               pr_info("oprofile: failed to allocate %d "
+                               "counters\n", num_counters);
+               ret = -ENOMEM;
+               num_counters = 0;
+               goto out;
+       }
+
+       for_each_possible_cpu(cpu) {
+               per_cpu(perf_events, cpu) = kcalloc(num_counters,
+                               sizeof(struct perf_event *), GFP_KERNEL);
+               if (!per_cpu(perf_events, cpu)) {
+                       pr_info("oprofile: failed to allocate %d perf events "
+                                       "for cpu %d\n", num_counters, cpu);
+                       ret = -ENOMEM;
+                       goto out;
+               }
+       }
+
+       ops->create_files       = oprofile_perf_create_files;
+       ops->setup              = oprofile_perf_setup;
+       ops->start              = oprofile_perf_start;
+       ops->stop               = oprofile_perf_stop;
+       ops->shutdown           = oprofile_perf_stop;
+       ops->cpu_type           = op_name_from_perf_id();
+
+       if (!ops->cpu_type)
+               ret = -ENODEV;
+       else
+               pr_info("oprofile: using %s\n", ops->cpu_type);
+
+out:
+       if (ret)
+               oprofile_perf_exit();
+
+       return ret;
+}