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