kernel/trace/trace_sched_wakeup.c

   1 /*
   2  * trace task wakeup timings
   3  *
   4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
   5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
   6  *
   7  * Based on code from the latency_tracer, that is:
   8  *
   9  *  Copyright (C) 2004-2006 Ingo Molnar
  10  *  Copyright (C) 2004 William Lee Irwin III
  11  */
  12 #include <linux/module.h>
  13 #include <linux/fs.h>
  14 #include <linux/debugfs.h>
  15 #include <linux/kallsyms.h>
  16 #include <linux/uaccess.h>
  17 #include <linux/ftrace.h>
  18 #include <trace/events/sched.h>
  19
  20 #include "trace.h"
  21
  22 static struct trace_array       *wakeup_trace;
  23 static int __read_mostly        tracer_enabled;
  24
  25 static struct task_struct       *wakeup_task;
  26 static int                      wakeup_cpu;
  27 static int                      wakeup_current_cpu;
  28 static unsigned                 wakeup_prio = -1;
  29 static int                      wakeup_rt;
  30
  31 static arch_spinlock_t wakeup_lock =
  32         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
  33
  34 static void __wakeup_reset(struct trace_array *tr);
  35
  36 static int save_lat_flag;
  37
  38 #ifdef CONFIG_FUNCTION_TRACER
  39 /*
  40  * irqsoff uses its own tracer function to keep the overhead down:
  41  */
  42 static void
  43 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
  44 {
  45         struct trace_array *tr = wakeup_trace;
  46         struct trace_array_cpu *data;
  47         unsigned long flags;
  48         long disabled;
  49         int cpu;
  50         int pc;
  51
  52         if (likely(!wakeup_task))
  53                 return;
  54
  55         pc = preempt_count();
  56         preempt_disable_notrace();
  57
  58         cpu = raw_smp_processor_id();
  59         if (cpu != wakeup_current_cpu)
  60                 goto out_enable;
  61
  62         data = tr->data[cpu];
  63         disabled = atomic_inc_return(&data->disabled);
  64         if (unlikely(disabled != 1))
  65                 goto out;
  66
  67         local_irq_save(flags);
  68
  69         trace_function(tr, ip, parent_ip, flags, pc);
  70
  71         local_irq_restore(flags);
  72
  73  out:
  74         atomic_dec(&data->disabled);
  75  out_enable:
  76         preempt_enable_notrace();
  77 }
  78
  79 static struct ftrace_ops trace_ops __read_mostly =
  80 {
  81         .func = wakeup_tracer_call,
  82 };
  83 #endif /* CONFIG_FUNCTION_TRACER */
  84
  85 /*
  86  * Should this new latency be reported/recorded?
  87  */
  88 static int report_latency(cycle_t delta)
  89 {
  90         if (tracing_thresh) {
  91                 if (delta < tracing_thresh)
  92                         return 0;
  93         } else {
  94                 if (delta <= tracing_max_latency)
  95                         return 0;
  96         }
  97         return 1;
  98 }
  99
 100 static void
 101 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
 102 {
 103         if (task != wakeup_task)
 104                 return;
 105
 106         wakeup_current_cpu = cpu;
 107 }
 108
 109 static void notrace
 110 probe_wakeup_sched_switch(void *ignore,
 111                           struct task_struct *prev, struct task_struct *next)
 112 {
 113         struct trace_array_cpu *data;
 114         cycle_t T0, T1, delta;
 115         unsigned long flags;
 116         long disabled;
 117         int cpu;
 118         int pc;
 119
 120         tracing_record_cmdline(prev);
 121
 122         if (unlikely(!tracer_enabled))
 123                 return;
 124
 125         /*
 126          * When we start a new trace, we set wakeup_task to NULL
 127          * and then set tracer_enabled = 1. We want to make sure
 128          * that another CPU does not see the tracer_enabled = 1
 129          * and the wakeup_task with an older task, that might
 130          * actually be the same as next.
 131          */
 132         smp_rmb();
 133
 134         if (next != wakeup_task)
 135                 return;
 136
 137         pc = preempt_count();
 138
 139         /* disable local data, not wakeup_cpu data */
 140         cpu = raw_smp_processor_id();
 141         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 142         if (likely(disabled != 1))
 143                 goto out;
 144
 145         local_irq_save(flags);
 146         arch_spin_lock(&wakeup_lock);
 147
 148         /* We could race with grabbing wakeup_lock */
 149         if (unlikely(!tracer_enabled || next != wakeup_task))
 150                 goto out_unlock;
 151
 152         /* The task we are waiting for is waking up */
 153         data = wakeup_trace->data[wakeup_cpu];
 154
 155         trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
 156         tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
 157
 158         T0 = data->preempt_timestamp;
 159         T1 = ftrace_now(cpu);
 160         delta = T1-T0;
 161
 162         if (!report_latency(delta))
 163                 goto out_unlock;
 164
 165         if (likely(!is_tracing_stopped())) {
 166                 tracing_max_latency = delta;
 167                 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
 168         }
 169
 170 out_unlock:
 171         __wakeup_reset(wakeup_trace);
 172         arch_spin_unlock(&wakeup_lock);
 173         local_irq_restore(flags);
 174 out:
 175         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 176 }
 177
 178 static void __wakeup_reset(struct trace_array *tr)
 179 {
 180         wakeup_cpu = -1;
 181         wakeup_prio = -1;
 182
 183         if (wakeup_task)
 184                 put_task_struct(wakeup_task);
 185
 186         wakeup_task = NULL;
 187 }
 188
 189 static void wakeup_reset(struct trace_array *tr)
 190 {
 191         unsigned long flags;
 192
 193         tracing_reset_online_cpus(tr);
 194
 195         local_irq_save(flags);
 196         arch_spin_lock(&wakeup_lock);
 197         __wakeup_reset(tr);
 198         arch_spin_unlock(&wakeup_lock);
 199         local_irq_restore(flags);
 200 }
 201
 202 static void
 203 probe_wakeup(void *ignore, struct task_struct *p, int success)
 204 {
 205         struct trace_array_cpu *data;
 206         int cpu = smp_processor_id();
 207         unsigned long flags;
 208         long disabled;
 209         int pc;
 210
 211         if (likely(!tracer_enabled))
 212                 return;
 213
 214         tracing_record_cmdline(p);
 215         tracing_record_cmdline(current);
 216
 217         if ((wakeup_rt && !rt_task(p)) ||
 218                         p->prio >= wakeup_prio ||
 219                         p->prio >= current->prio)
 220                 return;
 221
 222         pc = preempt_count();
 223         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 224         if (unlikely(disabled != 1))
 225                 goto out;
 226
 227         /* interrupts should be off from try_to_wake_up */
 228         arch_spin_lock(&wakeup_lock);
 229
 230         /* check for races. */
 231         if (!tracer_enabled || p->prio >= wakeup_prio)
 232                 goto out_locked;
 233
 234         /* reset the trace */
 235         __wakeup_reset(wakeup_trace);
 236
 237         wakeup_cpu = task_cpu(p);
 238         wakeup_current_cpu = wakeup_cpu;
 239         wakeup_prio = p->prio;
 240
 241         wakeup_task = p;
 242         get_task_struct(wakeup_task);
 243
 244         local_save_flags(flags);
 245
 246         data = wakeup_trace->data[wakeup_cpu];
 247         data->preempt_timestamp = ftrace_now(cpu);
 248         tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
 249
 250         /*
 251          * We must be careful in using CALLER_ADDR2. But since wake_up
 252          * is not called by an assembly function  (where as schedule is)
 253          * it should be safe to use it here.
 254          */
 255         trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
 256
 257 out_locked:
 258         arch_spin_unlock(&wakeup_lock);
 259 out:
 260         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 261 }
 262
 263 static void start_wakeup_tracer(struct trace_array *tr)
 264 {
 265         int ret;
 266
 267         ret = register_trace_sched_wakeup(probe_wakeup, NULL);
 268         if (ret) {
 269                 pr_info("wakeup trace: Couldn't activate tracepoint"
 270                         " probe to kernel_sched_wakeup\n");
 271                 return;
 272         }
 273
 274         ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
 275         if (ret) {
 276                 pr_info("wakeup trace: Couldn't activate tracepoint"
 277                         " probe to kernel_sched_wakeup_new\n");
 278                 goto fail_deprobe;
 279         }
 280
 281         ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
 282         if (ret) {
 283                 pr_info("sched trace: Couldn't activate tracepoint"
 284                         " probe to kernel_sched_switch\n");
 285                 goto fail_deprobe_wake_new;
 286         }
 287
 288         ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
 289         if (ret) {
 290                 pr_info("wakeup trace: Couldn't activate tracepoint"
 291                         " probe to kernel_sched_migrate_task\n");
 292                 return;
 293         }
 294
 295         wakeup_reset(tr);
 296
 297         /*
 298          * Don't let the tracer_enabled = 1 show up before
 299          * the wakeup_task is reset. This may be overkill since
 300          * wakeup_reset does a spin_unlock after setting the
 301          * wakeup_task to NULL, but I want to be safe.
 302          * This is a slow path anyway.
 303          */
 304         smp_wmb();
 305
 306         register_ftrace_function(&trace_ops);
 307
 308         if (tracing_is_enabled())
 309                 tracer_enabled = 1;
 310         else
 311                 tracer_enabled = 0;
 312
 313         return;
 314 fail_deprobe_wake_new:
 315         unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
 316 fail_deprobe:
 317         unregister_trace_sched_wakeup(probe_wakeup, NULL);
 318 }
 319
 320 static void stop_wakeup_tracer(struct trace_array *tr)
 321 {
 322         tracer_enabled = 0;
 323         unregister_ftrace_function(&trace_ops);
 324         unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
 325         unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
 326         unregister_trace_sched_wakeup(probe_wakeup, NULL);
 327         unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
 328 }
 329
 330 static int __wakeup_tracer_init(struct trace_array *tr)
 331 {
 332         save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
 333         trace_flags |= TRACE_ITER_LATENCY_FMT;
 334
 335         tracing_max_latency = 0;
 336         wakeup_trace = tr;
 337         start_wakeup_tracer(tr);
 338         return 0;
 339 }
 340
 341 static int wakeup_tracer_init(struct trace_array *tr)
 342 {
 343         wakeup_rt = 0;
 344         return __wakeup_tracer_init(tr);
 345 }
 346
 347 static int wakeup_rt_tracer_init(struct trace_array *tr)
 348 {
 349         wakeup_rt = 1;
 350         return __wakeup_tracer_init(tr);
 351 }
 352
 353 static void wakeup_tracer_reset(struct trace_array *tr)
 354 {
 355         stop_wakeup_tracer(tr);
 356         /* make sure we put back any tasks we are tracing */
 357         wakeup_reset(tr);
 358
 359         if (!save_lat_flag)
 360                 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
 361 }
 362
 363 static void wakeup_tracer_start(struct trace_array *tr)
 364 {
 365         wakeup_reset(tr);
 366         tracer_enabled = 1;
 367 }
 368
 369 static void wakeup_tracer_stop(struct trace_array *tr)
 370 {
 371         tracer_enabled = 0;
 372 }
 373
 374 static struct tracer wakeup_tracer __read_mostly =
 375 {
 376         .name           = "wakeup",
 377         .init           = wakeup_tracer_init,
 378         .reset          = wakeup_tracer_reset,
 379         .start          = wakeup_tracer_start,
 380         .stop           = wakeup_tracer_stop,
 381         .print_max      = 1,
 382 #ifdef CONFIG_FTRACE_SELFTEST
 383         .selftest    = trace_selftest_startup_wakeup,
 384 #endif
 385         .use_max_tr     = 1,
 386 };
 387
 388 static struct tracer wakeup_rt_tracer __read_mostly =
 389 {
 390         .name           = "wakeup_rt",
 391         .init           = wakeup_rt_tracer_init,
 392         .reset          = wakeup_tracer_reset,
 393         .start          = wakeup_tracer_start,
 394         .stop           = wakeup_tracer_stop,
 395         .wait_pipe      = poll_wait_pipe,
 396         .print_max      = 1,
 397 #ifdef CONFIG_FTRACE_SELFTEST
 398         .selftest    = trace_selftest_startup_wakeup,
 399 #endif
 400         .use_max_tr     = 1,
 401 };
 402
 403 __init static int init_wakeup_tracer(void)
 404 {
 405         int ret;
 406
 407         ret = register_tracer(&wakeup_tracer);
 408         if (ret)
 409                 return ret;
 410
 411         ret = register_tracer(&wakeup_rt_tracer);
 412         if (ret)
 413                 return ret;
 414
 415         return 0;
 416 }
 417 device_initcall(init_wakeup_tracer);