arch/arm/oprofile/common.c

   1 /**
   2  * @file common.c
   3  *
   4  * @remark Copyright 2004 Oprofile Authors
   5  * @remark Copyright 2010 ARM Ltd.
   6  * @remark Read the file COPYING
   7  *
   8  * @author Zwane Mwaikambo
   9  * @author Will Deacon [move to perf]
  10  */
  11
  12 #include <linux/cpumask.h>
  13 #include <linux/err.h>
  14 #include <linux/errno.h>
  15 #include <linux/init.h>
  16 #include <linux/mutex.h>
  17 #include <linux/oprofile.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/platform_device.h>
  20 #include <linux/slab.h>
  21 #include <asm/stacktrace.h>
  22 #include <linux/uaccess.h>
  23
  24 #include <asm/perf_event.h>
  25 #include <asm/ptrace.h>
  26
  27 #ifdef CONFIG_HW_PERF_EVENTS
  28 /*
  29  * Per performance monitor configuration as set via oprofilefs.
  30  */
  31 struct op_counter_config {
  32         unsigned long count;
  33         unsigned long enabled;
  34         unsigned long event;
  35         unsigned long unit_mask;
  36         unsigned long kernel;
  37         unsigned long user;
  38         struct perf_event_attr attr;
  39 };
  40
  41 static int op_arm_enabled;
  42 static DEFINE_MUTEX(op_arm_mutex);
  43
  44 static struct op_counter_config *counter_config;
  45 static struct perf_event **perf_events[nr_cpumask_bits];
  46 static int perf_num_counters;
  47
  48 /*
  49  * Overflow callback for oprofile.
  50  */
  51 static void op_overflow_handler(struct perf_event *event, int unused,
  52                         struct perf_sample_data *data, struct pt_regs *regs)
  53 {
  54         int id;
  55         u32 cpu = smp_processor_id();
  56
  57         for (id = 0; id < perf_num_counters; ++id)
  58                 if (perf_events[cpu][id] == event)
  59                         break;
  60
  61         if (id != perf_num_counters)
  62                 oprofile_add_sample(regs, id);
  63         else
  64                 pr_warning("oprofile: ignoring spurious overflow "
  65                                 "on cpu %u\n", cpu);
  66 }
  67
  68 /*
  69  * Called by op_arm_setup to create perf attributes to mirror the oprofile
  70  * settings in counter_config. Attributes are created as `pinned' events and
  71  * so are permanently scheduled on the PMU.
  72  */
  73 static void op_perf_setup(void)
  74 {
  75         int i;
  76         u32 size = sizeof(struct perf_event_attr);
  77         struct perf_event_attr *attr;
  78
  79         for (i = 0; i < perf_num_counters; ++i) {
  80                 attr = &counter_config[i].attr;
  81                 memset(attr, 0, size);
  82                 attr->type              = PERF_TYPE_RAW;
  83                 attr->size              = size;
  84                 attr->config            = counter_config[i].event;
  85                 attr->sample_period     = counter_config[i].count;
  86                 attr->pinned            = 1;
  87         }
  88 }
  89
  90 static int op_create_counter(int cpu, int event)
  91 {
  92         struct perf_event *pevent;
  93
  94         if (!counter_config[event].enabled || perf_events[cpu][event])
  95                 return 0;
  96
  97         pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
  98                                                   cpu, -1,
  99                                                   op_overflow_handler);
 100
 101         if (IS_ERR(pevent))
 102                 return PTR_ERR(pevent);
 103
 104         if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
 105                 perf_event_release_kernel(pevent);
 106                 pr_warning("oprofile: failed to enable event %d "
 107                                 "on CPU %d\n", event, cpu);
 108                 return -EBUSY;
 109         }
 110
 111         perf_events[cpu][event] = pevent;
 112
 113         return 0;
 114 }
 115
 116 static void op_destroy_counter(int cpu, int event)
 117 {
 118         struct perf_event *pevent = perf_events[cpu][event];
 119
 120         if (pevent) {
 121                 perf_event_release_kernel(pevent);
 122                 perf_events[cpu][event] = NULL;
 123         }
 124 }
 125
 126 /*
 127  * Called by op_arm_start to create active perf events based on the
 128  * perviously configured attributes.
 129  */
 130 static int op_perf_start(void)
 131 {
 132         int cpu, event, ret = 0;
 133
 134         for_each_online_cpu(cpu) {
 135                 for (event = 0; event < perf_num_counters; ++event) {
 136                         ret = op_create_counter(cpu, event);
 137                         if (ret)
 138                                 goto out;
 139                 }
 140         }
 141
 142 out:
 143         return ret;
 144 }
 145
 146 /*
 147  * Called by op_arm_stop at the end of a profiling run.
 148  */
 149 static void op_perf_stop(void)
 150 {
 151         int cpu, event;
 152
 153         for_each_online_cpu(cpu)
 154                 for (event = 0; event < perf_num_counters; ++event)
 155                         op_destroy_counter(cpu, event);
 156 }
 157
 158
 159 static char *op_name_from_perf_id(enum arm_perf_pmu_ids id)
 160 {
 161         switch (id) {
 162         case ARM_PERF_PMU_ID_XSCALE1:
 163                 return "arm/xscale1";
 164         case ARM_PERF_PMU_ID_XSCALE2:
 165                 return "arm/xscale2";
 166         case ARM_PERF_PMU_ID_V6:
 167                 return "arm/armv6";
 168         case ARM_PERF_PMU_ID_V6MP:
 169                 return "arm/mpcore";
 170         case ARM_PERF_PMU_ID_CA8:
 171                 return "arm/armv7";
 172         case ARM_PERF_PMU_ID_CA9:
 173                 return "arm/armv7-ca9";
 174         default:
 175                 return NULL;
 176         }
 177 }
 178
 179 static int op_arm_create_files(struct super_block *sb, struct dentry *root)
 180 {
 181         unsigned int i;
 182
 183         for (i = 0; i < perf_num_counters; i++) {
 184                 struct dentry *dir;
 185                 char buf[4];
 186
 187                 snprintf(buf, sizeof buf, "%d", i);
 188                 dir = oprofilefs_mkdir(sb, root, buf);
 189                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
 190                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
 191                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
 192                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
 193                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
 194                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
 195         }
 196
 197         return 0;
 198 }
 199
 200 static int op_arm_setup(void)
 201 {
 202         spin_lock(&oprofilefs_lock);
 203         op_perf_setup();
 204         spin_unlock(&oprofilefs_lock);
 205         return 0;
 206 }
 207
 208 static int op_arm_start(void)
 209 {
 210         int ret = -EBUSY;
 211
 212         mutex_lock(&op_arm_mutex);
 213         if (!op_arm_enabled) {
 214                 ret = 0;
 215                 op_perf_start();
 216                 op_arm_enabled = 1;
 217         }
 218         mutex_unlock(&op_arm_mutex);
 219         return ret;
 220 }
 221
 222 static void op_arm_stop(void)
 223 {
 224         mutex_lock(&op_arm_mutex);
 225         if (op_arm_enabled)
 226                 op_perf_stop();
 227         op_arm_enabled = 0;
 228         mutex_unlock(&op_arm_mutex);
 229 }
 230
 231 #ifdef CONFIG_PM
 232 static int op_arm_suspend(struct platform_device *dev, pm_message_t state)
 233 {
 234         mutex_lock(&op_arm_mutex);
 235         if (op_arm_enabled)
 236                 op_perf_stop();
 237         mutex_unlock(&op_arm_mutex);
 238         return 0;
 239 }
 240
 241 static int op_arm_resume(struct platform_device *dev)
 242 {
 243         mutex_lock(&op_arm_mutex);
 244         if (op_arm_enabled && op_perf_start())
 245                 op_arm_enabled = 0;
 246         mutex_unlock(&op_arm_mutex);
 247         return 0;
 248 }
 249
 250 static struct platform_driver oprofile_driver = {
 251         .driver         = {
 252                 .name           = "arm-oprofile",
 253         },
 254         .resume         = op_arm_resume,
 255         .suspend        = op_arm_suspend,
 256 };
 257
 258 static struct platform_device *oprofile_pdev;
 259
 260 static int __init init_driverfs(void)
 261 {
 262         int ret;
 263
 264         ret = platform_driver_register(&oprofile_driver);
 265         if (ret)
 266                 goto out;
 267
 268         oprofile_pdev = platform_device_register_simple(
 269                                 oprofile_driver.driver.name, 0, NULL, 0);
 270         if (IS_ERR(oprofile_pdev)) {
 271                 ret = PTR_ERR(oprofile_pdev);
 272                 platform_driver_unregister(&oprofile_driver);
 273         }
 274
 275 out:
 276         return ret;
 277 }
 278
 279 static void __exit exit_driverfs(void)
 280 {
 281         platform_device_unregister(oprofile_pdev);
 282         platform_driver_unregister(&oprofile_driver);
 283 }
 284 #else
 285 static int __init init_driverfs(void) { return 0; }
 286 #define exit_driverfs() do { } while (0)
 287 #endif /* CONFIG_PM */
 288
 289 static int report_trace(struct stackframe *frame, void *d)
 290 {
 291         unsigned int *depth = d;
 292
 293         if (*depth) {
 294                 oprofile_add_trace(frame->pc);
 295                 (*depth)--;
 296         }
 297
 298         return *depth == 0;
 299 }
 300
 301 /*
 302  * The registers we're interested in are at the end of the variable
 303  * length saved register structure. The fp points at the end of this
 304  * structure so the address of this struct is:
 305  * (struct frame_tail *)(xxx->fp)-1
 306  */
 307 struct frame_tail {
 308         struct frame_tail *fp;
 309         unsigned long sp;
 310         unsigned long lr;
 311 } __attribute__((packed));
 312
 313 static struct frame_tail* user_backtrace(struct frame_tail *tail)
 314 {
 315         struct frame_tail buftail[2];
 316
 317         /* Also check accessibility of one struct frame_tail beyond */
 318         if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
 319                 return NULL;
 320         if (__copy_from_user_inatomic(buftail, tail, sizeof(buftail)))
 321                 return NULL;
 322
 323         oprofile_add_trace(buftail[0].lr);
 324
 325         /* frame pointers should strictly progress back up the stack
 326          * (towards higher addresses) */
 327         if (tail >= buftail[0].fp)
 328                 return NULL;
 329
 330         return buftail[0].fp-1;
 331 }
 332
 333 static void arm_backtrace(struct pt_regs * const regs, unsigned int depth)
 334 {
 335         struct frame_tail *tail = ((struct frame_tail *) regs->ARM_fp) - 1;
 336
 337         if (!user_mode(regs)) {
 338                 struct stackframe frame;
 339                 frame.fp = regs->ARM_fp;
 340                 frame.sp = regs->ARM_sp;
 341                 frame.lr = regs->ARM_lr;
 342                 frame.pc = regs->ARM_pc;
 343                 walk_stackframe(&frame, report_trace, &depth);
 344                 return;
 345         }
 346
 347         while (depth-- && tail && !((unsigned long) tail & 3))
 348                 tail = user_backtrace(tail);
 349 }
 350
 351 void oprofile_arch_exit(void)
 352 {
 353         int cpu, id;
 354         struct perf_event *event;
 355
 356         for_each_possible_cpu(cpu) {
 357                 for (id = 0; id < perf_num_counters; ++id) {
 358                         event = perf_events[cpu][id];
 359                         if (event)
 360                                 perf_event_release_kernel(event);
 361                 }
 362
 363                 kfree(perf_events[cpu]);
 364         }
 365
 366         kfree(counter_config);
 367         exit_driverfs();
 368 }
 369
 370 int __init oprofile_arch_init(struct oprofile_operations *ops)
 371 {
 372         int cpu, ret = 0;
 373
 374         ret = init_driverfs();
 375         if (ret)
 376                 return ret;
 377
 378         memset(&perf_events, 0, sizeof(perf_events));
 379
 380         perf_num_counters = armpmu_get_max_events();
 381
 382         counter_config = kcalloc(perf_num_counters,
 383                         sizeof(struct op_counter_config), GFP_KERNEL);
 384
 385         if (!counter_config) {
 386                 pr_info("oprofile: failed to allocate %d "
 387                                 "counters\n", perf_num_counters);
 388                 ret = -ENOMEM;
 389                 perf_num_counters = 0;
 390                 goto out;
 391         }
 392
 393         for_each_possible_cpu(cpu) {
 394                 perf_events[cpu] = kcalloc(perf_num_counters,
 395                                 sizeof(struct perf_event *), GFP_KERNEL);
 396                 if (!perf_events[cpu]) {
 397                         pr_info("oprofile: failed to allocate %d perf events "
 398                                         "for cpu %d\n", perf_num_counters, cpu);
 399                         ret = -ENOMEM;
 400                         goto out;
 401                 }
 402         }
 403
 404         ops->backtrace          = arm_backtrace;
 405         ops->create_files       = op_arm_create_files;
 406         ops->setup              = op_arm_setup;
 407         ops->start              = op_arm_start;
 408         ops->stop               = op_arm_stop;
 409         ops->shutdown           = op_arm_stop;
 410         ops->cpu_type           = op_name_from_perf_id(armpmu_get_pmu_id());
 411
 412         if (!ops->cpu_type)
 413                 ret = -ENODEV;
 414         else
 415                 pr_info("oprofile: using %s\n", ops->cpu_type);
 416
 417 out:
 418         if (ret)
 419                 oprofile_arch_exit();
 420
 421         return ret;
 422 }
 423
 424 #else
 425 int __init oprofile_arch_init(struct oprofile_operations *ops)
 426 {
 427         pr_info("oprofile: hardware counters not available\n");
 428         return -ENODEV;
 429 }
 430 void __exit oprofile_arch_exit(void) {}
 431 #endif /* CONFIG_HW_PERF_EVENTS */