drivers/oprofile/oprofile_perf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2010 ARM Ltd.
   4  * Copyright 2012 Advanced Micro Devices, Inc., Robert Richter
   5  *
   6  * Perf-events backend for OProfile.
   7  */
   8 #include <linux/perf_event.h>
   9 #include <linux/platform_device.h>
  10 #include <linux/oprofile.h>
  11 #include <linux/slab.h>
  12
  13 /*
  14  * Per performance monitor configuration as set via oprofilefs.
  15  */
  16 struct op_counter_config {
  17         unsigned long count;
  18         unsigned long enabled;
  19         unsigned long event;
  20         unsigned long unit_mask;
  21         unsigned long kernel;
  22         unsigned long user;
  23         struct perf_event_attr attr;
  24 };
  25
  26 static int oprofile_perf_enabled;
  27 static DEFINE_MUTEX(oprofile_perf_mutex);
  28
  29 static struct op_counter_config *counter_config;
  30 static DEFINE_PER_CPU(struct perf_event **, perf_events);
  31 static int num_counters;
  32
  33 /*
  34  * Overflow callback for oprofile.
  35  */
  36 static void op_overflow_handler(struct perf_event *event,
  37                         struct perf_sample_data *data, struct pt_regs *regs)
  38 {
  39         int id;
  40         u32 cpu = smp_processor_id();
  41
  42         for (id = 0; id < num_counters; ++id)
  43                 if (per_cpu(perf_events, cpu)[id] == event)
  44                         break;
  45
  46         if (id != num_counters)
  47                 oprofile_add_sample(regs, id);
  48         else
  49                 pr_warning("oprofile: ignoring spurious overflow "
  50                                 "on cpu %u\n", cpu);
  51 }
  52
  53 /*
  54  * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile
  55  * settings in counter_config. Attributes are created as `pinned' events and
  56  * so are permanently scheduled on the PMU.
  57  */
  58 static void op_perf_setup(void)
  59 {
  60         int i;
  61         u32 size = sizeof(struct perf_event_attr);
  62         struct perf_event_attr *attr;
  63
  64         for (i = 0; i < num_counters; ++i) {
  65                 attr = &counter_config[i].attr;
  66                 memset(attr, 0, size);
  67                 attr->type              = PERF_TYPE_RAW;
  68                 attr->size              = size;
  69                 attr->config            = counter_config[i].event;
  70                 attr->sample_period     = counter_config[i].count;
  71                 attr->pinned            = 1;
  72         }
  73 }
  74
  75 static int op_create_counter(int cpu, int event)
  76 {
  77         struct perf_event *pevent;
  78
  79         if (!counter_config[event].enabled || per_cpu(perf_events, cpu)[event])
  80                 return 0;
  81
  82         pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
  83                                                   cpu, NULL,
  84                                                   op_overflow_handler, NULL);
  85
  86         if (IS_ERR(pevent))
  87                 return PTR_ERR(pevent);
  88
  89         if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
  90                 perf_event_release_kernel(pevent);
  91                 pr_warning("oprofile: failed to enable event %d "
  92                                 "on CPU %d\n", event, cpu);
  93                 return -EBUSY;
  94         }
  95
  96         per_cpu(perf_events, cpu)[event] = pevent;
  97
  98         return 0;
  99 }
 100
 101 static void op_destroy_counter(int cpu, int event)
 102 {
 103         struct perf_event *pevent = per_cpu(perf_events, cpu)[event];
 104
 105         if (pevent) {
 106                 perf_event_release_kernel(pevent);
 107                 per_cpu(perf_events, cpu)[event] = NULL;
 108         }
 109 }
 110
 111 /*
 112  * Called by oprofile_perf_start to create active perf events based on the
 113  * perviously configured attributes.
 114  */
 115 static int op_perf_start(void)
 116 {
 117         int cpu, event, ret = 0;
 118
 119         for_each_online_cpu(cpu) {
 120                 for (event = 0; event < num_counters; ++event) {
 121                         ret = op_create_counter(cpu, event);
 122                         if (ret)
 123                                 return ret;
 124                 }
 125         }
 126
 127         return ret;
 128 }
 129
 130 /*
 131  * Called by oprofile_perf_stop at the end of a profiling run.
 132  */
 133 static void op_perf_stop(void)
 134 {
 135         int cpu, event;
 136
 137         for_each_online_cpu(cpu)
 138                 for (event = 0; event < num_counters; ++event)
 139                         op_destroy_counter(cpu, event);
 140 }
 141
 142 static int oprofile_perf_create_files(struct dentry *root)
 143 {
 144         unsigned int i;
 145
 146         for (i = 0; i < num_counters; i++) {
 147                 struct dentry *dir;
 148                 char buf[4];
 149
 150                 snprintf(buf, sizeof buf, "%d", i);
 151                 dir = oprofilefs_mkdir(root, buf);
 152                 oprofilefs_create_ulong(dir, "enabled", &counter_config[i].enabled);
 153                 oprofilefs_create_ulong(dir, "event", &counter_config[i].event);
 154                 oprofilefs_create_ulong(dir, "count", &counter_config[i].count);
 155                 oprofilefs_create_ulong(dir, "unit_mask", &counter_config[i].unit_mask);
 156                 oprofilefs_create_ulong(dir, "kernel", &counter_config[i].kernel);
 157                 oprofilefs_create_ulong(dir, "user", &counter_config[i].user);
 158         }
 159
 160         return 0;
 161 }
 162
 163 static int oprofile_perf_setup(void)
 164 {
 165         raw_spin_lock(&oprofilefs_lock);
 166         op_perf_setup();
 167         raw_spin_unlock(&oprofilefs_lock);
 168         return 0;
 169 }
 170
 171 static int oprofile_perf_start(void)
 172 {
 173         int ret = -EBUSY;
 174
 175         mutex_lock(&oprofile_perf_mutex);
 176         if (!oprofile_perf_enabled) {
 177                 ret = 0;
 178                 op_perf_start();
 179                 oprofile_perf_enabled = 1;
 180         }
 181         mutex_unlock(&oprofile_perf_mutex);
 182         return ret;
 183 }
 184
 185 static void oprofile_perf_stop(void)
 186 {
 187         mutex_lock(&oprofile_perf_mutex);
 188         if (oprofile_perf_enabled)
 189                 op_perf_stop();
 190         oprofile_perf_enabled = 0;
 191         mutex_unlock(&oprofile_perf_mutex);
 192 }
 193
 194 #ifdef CONFIG_PM
 195
 196 static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state)
 197 {
 198         mutex_lock(&oprofile_perf_mutex);
 199         if (oprofile_perf_enabled)
 200                 op_perf_stop();
 201         mutex_unlock(&oprofile_perf_mutex);
 202         return 0;
 203 }
 204
 205 static int oprofile_perf_resume(struct platform_device *dev)
 206 {
 207         mutex_lock(&oprofile_perf_mutex);
 208         if (oprofile_perf_enabled && op_perf_start())
 209                 oprofile_perf_enabled = 0;
 210         mutex_unlock(&oprofile_perf_mutex);
 211         return 0;
 212 }
 213
 214 static struct platform_driver oprofile_driver = {
 215         .driver         = {
 216                 .name           = "oprofile-perf",
 217         },
 218         .resume         = oprofile_perf_resume,
 219         .suspend        = oprofile_perf_suspend,
 220 };
 221
 222 static struct platform_device *oprofile_pdev;
 223
 224 static int __init init_driverfs(void)
 225 {
 226         int ret;
 227
 228         ret = platform_driver_register(&oprofile_driver);
 229         if (ret)
 230                 return ret;
 231
 232         oprofile_pdev = platform_device_register_simple(
 233                                 oprofile_driver.driver.name, 0, NULL, 0);
 234         if (IS_ERR(oprofile_pdev)) {
 235                 ret = PTR_ERR(oprofile_pdev);
 236                 platform_driver_unregister(&oprofile_driver);
 237         }
 238
 239         return ret;
 240 }
 241
 242 static void exit_driverfs(void)
 243 {
 244         platform_device_unregister(oprofile_pdev);
 245         platform_driver_unregister(&oprofile_driver);
 246 }
 247
 248 #else
 249
 250 static inline int  init_driverfs(void) { return 0; }
 251 static inline void exit_driverfs(void) { }
 252
 253 #endif /* CONFIG_PM */
 254
 255 void oprofile_perf_exit(void)
 256 {
 257         int cpu, id;
 258         struct perf_event *event;
 259
 260         for_each_possible_cpu(cpu) {
 261                 for (id = 0; id < num_counters; ++id) {
 262                         event = per_cpu(perf_events, cpu)[id];
 263                         if (event)
 264                                 perf_event_release_kernel(event);
 265                 }
 266
 267                 kfree(per_cpu(perf_events, cpu));
 268         }
 269
 270         kfree(counter_config);
 271         exit_driverfs();
 272 }
 273
 274 int __init oprofile_perf_init(struct oprofile_operations *ops)
 275 {
 276         int cpu, ret = 0;
 277
 278         ret = init_driverfs();
 279         if (ret)
 280                 return ret;
 281
 282         num_counters = perf_num_counters();
 283         if (num_counters <= 0) {
 284                 pr_info("oprofile: no performance counters\n");
 285                 ret = -ENODEV;
 286                 goto out;
 287         }
 288
 289         counter_config = kcalloc(num_counters,
 290                         sizeof(struct op_counter_config), GFP_KERNEL);
 291
 292         if (!counter_config) {
 293                 pr_info("oprofile: failed to allocate %d "
 294                                 "counters\n", num_counters);
 295                 ret = -ENOMEM;
 296                 num_counters = 0;
 297                 goto out;
 298         }
 299
 300         for_each_possible_cpu(cpu) {
 301                 per_cpu(perf_events, cpu) = kcalloc(num_counters,
 302                                 sizeof(struct perf_event *), GFP_KERNEL);
 303                 if (!per_cpu(perf_events, cpu)) {
 304                         pr_info("oprofile: failed to allocate %d perf events "
 305                                         "for cpu %d\n", num_counters, cpu);
 306                         ret = -ENOMEM;
 307                         goto out;
 308                 }
 309         }
 310
 311         ops->create_files       = oprofile_perf_create_files;
 312         ops->setup              = oprofile_perf_setup;
 313         ops->start              = oprofile_perf_start;
 314         ops->stop               = oprofile_perf_stop;
 315         ops->shutdown           = oprofile_perf_stop;
 316         ops->cpu_type           = op_name_from_perf_id();
 317
 318         if (!ops->cpu_type)
 319                 ret = -ENODEV;
 320         else
 321                 pr_info("oprofile: using %s\n", ops->cpu_type);
 322
 323 out:
 324         if (ret)
 325                 oprofile_perf_exit();
 326
 327         return ret;
 328 }