drivers/oprofile/oprofile_perf.c

   1 /*
   2  * Copyright 2010 ARM Ltd.
   3  *
   4  * Perf-events backend for OProfile.
   5  */
   6 #include <linux/perf_event.h>
   7 #include <linux/platform_device.h>
   8 #include <linux/oprofile.h>
   9 #include <linux/slab.h>
  10
  11 /*
  12  * Per performance monitor configuration as set via oprofilefs.
  13  */
  14 struct op_counter_config {
  15         unsigned long count;
  16         unsigned long enabled;
  17         unsigned long event;
  18         unsigned long unit_mask;
  19         unsigned long kernel;
  20         unsigned long user;
  21         struct perf_event_attr attr;
  22 };
  23
  24 static int oprofile_perf_enabled;
  25 static DEFINE_MUTEX(oprofile_perf_mutex);
  26
  27 static struct op_counter_config *counter_config;
  28 static struct perf_event **perf_events[nr_cpumask_bits];
  29 static int num_counters;
  30
  31 /*
  32  * Overflow callback for oprofile.
  33  */
  34 static void op_overflow_handler(struct perf_event *event, int unused,
  35                         struct perf_sample_data *data, struct pt_regs *regs)
  36 {
  37         int id;
  38         u32 cpu = smp_processor_id();
  39
  40         for (id = 0; id < num_counters; ++id)
  41                 if (perf_events[cpu][id] == event)
  42                         break;
  43
  44         if (id != num_counters)
  45                 oprofile_add_sample(regs, id);
  46         else
  47                 pr_warning("oprofile: ignoring spurious overflow "
  48                                 "on cpu %u\n", cpu);
  49 }
  50
  51 /*
  52  * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile
  53  * settings in counter_config. Attributes are created as `pinned' events and
  54  * so are permanently scheduled on the PMU.
  55  */
  56 static void op_perf_setup(void)
  57 {
  58         int i;
  59         u32 size = sizeof(struct perf_event_attr);
  60         struct perf_event_attr *attr;
  61
  62         for (i = 0; i < num_counters; ++i) {
  63                 attr = &counter_config[i].attr;
  64                 memset(attr, 0, size);
  65                 attr->type              = PERF_TYPE_RAW;
  66                 attr->size              = size;
  67                 attr->config            = counter_config[i].event;
  68                 attr->sample_period     = counter_config[i].count;
  69                 attr->pinned            = 1;
  70         }
  71 }
  72
  73 static int op_create_counter(int cpu, int event)
  74 {
  75         struct perf_event *pevent;
  76
  77         if (!counter_config[event].enabled || perf_events[cpu][event])
  78                 return 0;
  79
  80         pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
  81                                                   cpu, NULL,
  82                                                   op_overflow_handler);
  83
  84         if (IS_ERR(pevent))
  85                 return PTR_ERR(pevent);
  86
  87         if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
  88                 perf_event_release_kernel(pevent);
  89                 pr_warning("oprofile: failed to enable event %d "
  90                                 "on CPU %d\n", event, cpu);
  91                 return -EBUSY;
  92         }
  93
  94         perf_events[cpu][event] = pevent;
  95
  96         return 0;
  97 }
  98
  99 static void op_destroy_counter(int cpu, int event)
 100 {
 101         struct perf_event *pevent = perf_events[cpu][event];
 102
 103         if (pevent) {
 104                 perf_event_release_kernel(pevent);
 105                 perf_events[cpu][event] = NULL;
 106         }
 107 }
 108
 109 /*
 110  * Called by oprofile_perf_start to create active perf events based on the
 111  * perviously configured attributes.
 112  */
 113 static int op_perf_start(void)
 114 {
 115         int cpu, event, ret = 0;
 116
 117         for_each_online_cpu(cpu) {
 118                 for (event = 0; event < num_counters; ++event) {
 119                         ret = op_create_counter(cpu, event);
 120                         if (ret)
 121                                 return ret;
 122                 }
 123         }
 124
 125         return ret;
 126 }
 127
 128 /*
 129  * Called by oprofile_perf_stop at the end of a profiling run.
 130  */
 131 static void op_perf_stop(void)
 132 {
 133         int cpu, event;
 134
 135         for_each_online_cpu(cpu)
 136                 for (event = 0; event < num_counters; ++event)
 137                         op_destroy_counter(cpu, event);
 138 }
 139
 140 static int oprofile_perf_create_files(struct super_block *sb, struct dentry *root)
 141 {
 142         unsigned int i;
 143
 144         for (i = 0; i < num_counters; i++) {
 145                 struct dentry *dir;
 146                 char buf[4];
 147
 148                 snprintf(buf, sizeof buf, "%d", i);
 149                 dir = oprofilefs_mkdir(sb, root, buf);
 150                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
 151                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
 152                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
 153                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
 154                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
 155                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
 156         }
 157
 158         return 0;
 159 }
 160
 161 static int oprofile_perf_setup(void)
 162 {
 163         spin_lock(&oprofilefs_lock);
 164         op_perf_setup();
 165         spin_unlock(&oprofilefs_lock);
 166         return 0;
 167 }
 168
 169 static int oprofile_perf_start(void)
 170 {
 171         int ret = -EBUSY;
 172
 173         mutex_lock(&oprofile_perf_mutex);
 174         if (!oprofile_perf_enabled) {
 175                 ret = 0;
 176                 op_perf_start();
 177                 oprofile_perf_enabled = 1;
 178         }
 179         mutex_unlock(&oprofile_perf_mutex);
 180         return ret;
 181 }
 182
 183 static void oprofile_perf_stop(void)
 184 {
 185         mutex_lock(&oprofile_perf_mutex);
 186         if (oprofile_perf_enabled)
 187                 op_perf_stop();
 188         oprofile_perf_enabled = 0;
 189         mutex_unlock(&oprofile_perf_mutex);
 190 }
 191
 192 #ifdef CONFIG_PM
 193 static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state)
 194 {
 195         mutex_lock(&oprofile_perf_mutex);
 196         if (oprofile_perf_enabled)
 197                 op_perf_stop();
 198         mutex_unlock(&oprofile_perf_mutex);
 199         return 0;
 200 }
 201
 202 static int oprofile_perf_resume(struct platform_device *dev)
 203 {
 204         mutex_lock(&oprofile_perf_mutex);
 205         if (oprofile_perf_enabled && op_perf_start())
 206                 oprofile_perf_enabled = 0;
 207         mutex_unlock(&oprofile_perf_mutex);
 208         return 0;
 209 }
 210
 211 static struct platform_driver oprofile_driver = {
 212         .driver         = {
 213                 .name           = "oprofile-perf",
 214         },
 215         .resume         = oprofile_perf_resume,
 216         .suspend        = oprofile_perf_suspend,
 217 };
 218
 219 static struct platform_device *oprofile_pdev;
 220
 221 static int __init init_driverfs(void)
 222 {
 223         int ret;
 224
 225         ret = platform_driver_register(&oprofile_driver);
 226         if (ret)
 227                 return ret;
 228
 229         oprofile_pdev = platform_device_register_simple(
 230                                 oprofile_driver.driver.name, 0, NULL, 0);
 231         if (IS_ERR(oprofile_pdev)) {
 232                 ret = PTR_ERR(oprofile_pdev);
 233                 platform_driver_unregister(&oprofile_driver);
 234         }
 235
 236         return ret;
 237 }
 238
 239 static void exit_driverfs(void)
 240 {
 241         platform_device_unregister(oprofile_pdev);
 242         platform_driver_unregister(&oprofile_driver);
 243 }
 244 #else
 245 static int __init init_driverfs(void) { return 0; }
 246 #define exit_driverfs() do { } while (0)
 247 #endif /* CONFIG_PM */
 248
 249 void oprofile_perf_exit(void)
 250 {
 251         int cpu, id;
 252         struct perf_event *event;
 253
 254         for_each_possible_cpu(cpu) {
 255                 for (id = 0; id < num_counters; ++id) {
 256                         event = perf_events[cpu][id];
 257                         if (event)
 258                                 perf_event_release_kernel(event);
 259                 }
 260
 261                 kfree(perf_events[cpu]);
 262         }
 263
 264         kfree(counter_config);
 265         exit_driverfs();
 266 }
 267
 268 int __init oprofile_perf_init(struct oprofile_operations *ops)
 269 {
 270         int cpu, ret = 0;
 271
 272         ret = init_driverfs();
 273         if (ret)
 274                 return ret;
 275
 276         memset(&perf_events, 0, sizeof(perf_events));
 277
 278         num_counters = perf_num_counters();
 279         if (num_counters <= 0) {
 280                 pr_info("oprofile: no performance counters\n");
 281                 ret = -ENODEV;
 282                 goto out;
 283         }
 284
 285         counter_config = kcalloc(num_counters,
 286                         sizeof(struct op_counter_config), GFP_KERNEL);
 287
 288         if (!counter_config) {
 289                 pr_info("oprofile: failed to allocate %d "
 290                                 "counters\n", num_counters);
 291                 ret = -ENOMEM;
 292                 num_counters = 0;
 293                 goto out;
 294         }
 295
 296         for_each_possible_cpu(cpu) {
 297                 perf_events[cpu] = kcalloc(num_counters,
 298                                 sizeof(struct perf_event *), GFP_KERNEL);
 299                 if (!perf_events[cpu]) {
 300                         pr_info("oprofile: failed to allocate %d perf events "
 301                                         "for cpu %d\n", num_counters, cpu);
 302                         ret = -ENOMEM;
 303                         goto out;
 304                 }
 305         }
 306
 307         ops->create_files       = oprofile_perf_create_files;
 308         ops->setup              = oprofile_perf_setup;
 309         ops->start              = oprofile_perf_start;
 310         ops->stop               = oprofile_perf_stop;
 311         ops->shutdown           = oprofile_perf_stop;
 312         ops->cpu_type           = op_name_from_perf_id();
 313
 314         if (!ops->cpu_type)
 315                 ret = -ENODEV;
 316         else
 317                 pr_info("oprofile: using %s\n", ops->cpu_type);
 318
 319 out:
 320         if (ret)
 321                 oprofile_perf_exit();
 322
 323         return ret;
 324 }