SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/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 unsigned                 wakeup_prio = -1;
  28 static int                      wakeup_rt;
  29
  30 static raw_spinlock_t wakeup_lock =
  31         (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  32
  33 static void __wakeup_reset(struct trace_array *tr);
  34
  35 #ifdef CONFIG_FUNCTION_TRACER
  36 /*
  37  * irqsoff uses its own tracer function to keep the overhead down:
  38  */
  39 static void
  40 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
  41 {
  42         struct trace_array *tr = wakeup_trace;
  43         struct trace_array_cpu *data;
  44         unsigned long flags;
  45         long disabled;
  46         int resched;
  47         int cpu;
  48         int pc;
  49
  50         if (likely(!wakeup_task))
  51                 return;
  52
  53         pc = preempt_count();
  54         resched = ftrace_preempt_disable();
  55
  56         cpu = raw_smp_processor_id();
  57         data = tr->data[cpu];
  58         disabled = atomic_inc_return(&data->disabled);
  59         if (unlikely(disabled != 1))
  60                 goto out;
  61
  62         local_irq_save(flags);
  63         __raw_spin_lock(&wakeup_lock);
  64
  65         if (unlikely(!wakeup_task))
  66                 goto unlock;
  67
  68         /*
  69          * The task can't disappear because it needs to
  70          * wake up first, and we have the wakeup_lock.
  71          */
  72         if (task_cpu(wakeup_task) != cpu)
  73                 goto unlock;
  74
  75         trace_function(tr, data, ip, parent_ip, flags, pc);
  76
  77  unlock:
  78         __raw_spin_unlock(&wakeup_lock);
  79         local_irq_restore(flags);
  80
  81  out:
  82         atomic_dec(&data->disabled);
  83
  84         ftrace_preempt_enable(resched);
  85 }
  86
  87 static struct ftrace_ops trace_ops __read_mostly =
  88 {
  89         .func = wakeup_tracer_call,
  90 };
  91 #endif /* CONFIG_FUNCTION_TRACER */
  92
  93 /*
  94  * Should this new latency be reported/recorded?
  95  */
  96 static int report_latency(cycle_t delta)
  97 {
  98         if (tracing_thresh) {
  99                 if (delta < tracing_thresh)
 100                         return 0;
 101         } else {
 102                 if (delta <= tracing_max_latency)
 103                         return 0;
 104         }
 105         return 1;
 106 }
 107
 108 static void notrace
 109 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
 110         struct task_struct *next)
 111 {
 112         unsigned long latency = 0, t0 = 0, t1 = 0;
 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         /* The task we are waiting for is waking up */
 140         data = wakeup_trace->data[wakeup_cpu];
 141
 142         /* disable local data, not wakeup_cpu data */
 143         cpu = raw_smp_processor_id();
 144         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 145         if (likely(disabled != 1))
 146                 goto out;
 147
 148         local_irq_save(flags);
 149         __raw_spin_lock(&wakeup_lock);
 150
 151         /* We could race with grabbing wakeup_lock */
 152         if (unlikely(!tracer_enabled || next != wakeup_task))
 153                 goto out_unlock;
 154
 155         trace_function(wakeup_trace, data, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
 156         tracing_sched_switch_trace(wakeup_trace, data, prev, next, flags, pc);
 157
 158         /*
 159          * usecs conversion is slow so we try to delay the conversion
 160          * as long as possible:
 161          */
 162         T0 = data->preempt_timestamp;
 163         T1 = ftrace_now(cpu);
 164         delta = T1-T0;
 165
 166         if (!report_latency(delta))
 167                 goto out_unlock;
 168
 169         latency = nsecs_to_usecs(delta);
 170
 171         tracing_max_latency = delta;
 172         t0 = nsecs_to_usecs(T0);
 173         t1 = nsecs_to_usecs(T1);
 174
 175         update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
 176
 177 out_unlock:
 178         __wakeup_reset(wakeup_trace);
 179         __raw_spin_unlock(&wakeup_lock);
 180         local_irq_restore(flags);
 181 out:
 182         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 183 }
 184
 185 static void __wakeup_reset(struct trace_array *tr)
 186 {
 187         struct trace_array_cpu *data;
 188         int cpu;
 189
 190         for_each_possible_cpu(cpu) {
 191                 data = tr->data[cpu];
 192                 tracing_reset(tr, cpu);
 193         }
 194
 195         wakeup_cpu = -1;
 196         wakeup_prio = -1;
 197
 198         if (wakeup_task)
 199                 put_task_struct(wakeup_task);
 200
 201         wakeup_task = NULL;
 202 }
 203
 204 static void wakeup_reset(struct trace_array *tr)
 205 {
 206         unsigned long flags;
 207
 208         local_irq_save(flags);
 209         __raw_spin_lock(&wakeup_lock);
 210         __wakeup_reset(tr);
 211         __raw_spin_unlock(&wakeup_lock);
 212         local_irq_restore(flags);
 213 }
 214
 215 static void
 216 probe_wakeup(struct rq *rq, struct task_struct *p, int success)
 217 {
 218         struct trace_array_cpu *data;
 219         int cpu = smp_processor_id();
 220         unsigned long flags;
 221         long disabled;
 222         int pc;
 223
 224         if (likely(!tracer_enabled))
 225                 return;
 226
 227         tracing_record_cmdline(p);
 228         tracing_record_cmdline(current);
 229
 230         if ((wakeup_rt && !rt_task(p)) ||
 231                         p->prio >= wakeup_prio ||
 232                         p->prio >= current->prio)
 233                 return;
 234
 235         pc = preempt_count();
 236         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 237         if (unlikely(disabled != 1))
 238                 goto out;
 239
 240         /* interrupts should be off from try_to_wake_up */
 241         __raw_spin_lock(&wakeup_lock);
 242
 243         /* check for races. */
 244         if (!tracer_enabled || p->prio >= wakeup_prio)
 245                 goto out_locked;
 246
 247         /* reset the trace */
 248         __wakeup_reset(wakeup_trace);
 249
 250         wakeup_cpu = task_cpu(p);
 251         wakeup_prio = p->prio;
 252
 253         wakeup_task = p;
 254         get_task_struct(wakeup_task);
 255
 256         local_save_flags(flags);
 257
 258         data = wakeup_trace->data[wakeup_cpu];
 259         data->preempt_timestamp = ftrace_now(cpu);
 260         tracing_sched_wakeup_trace(wakeup_trace, data, p, current,
 261                                    flags, pc);
 262         trace_function(wakeup_trace, data, CALLER_ADDR1, CALLER_ADDR2,
 263                        flags, pc);
 264
 265 out_locked:
 266         __raw_spin_unlock(&wakeup_lock);
 267 out:
 268         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 269 }
 270
 271 static void start_wakeup_tracer(struct trace_array *tr)
 272 {
 273         int ret;
 274
 275         ret = register_trace_sched_wakeup(probe_wakeup);
 276         if (ret) {
 277                 pr_info("wakeup trace: Couldn't activate tracepoint"
 278                         " probe to kernel_sched_wakeup\n");
 279                 return;
 280         }
 281
 282         ret = register_trace_sched_wakeup_new(probe_wakeup);
 283         if (ret) {
 284                 pr_info("wakeup trace: Couldn't activate tracepoint"
 285                         " probe to kernel_sched_wakeup_new\n");
 286                 goto fail_deprobe;
 287         }
 288
 289         ret = register_trace_sched_switch(probe_wakeup_sched_switch);
 290         if (ret) {
 291                 pr_info("sched trace: Couldn't activate tracepoint"
 292                         " probe to kernel_sched_schedule\n");
 293                 goto fail_deprobe_wake_new;
 294         }
 295
 296         wakeup_reset(tr);
 297
 298         /*
 299          * Don't let the tracer_enabled = 1 show up before
 300          * the wakeup_task is reset. This may be overkill since
 301          * wakeup_reset does a spin_unlock after setting the
 302          * wakeup_task to NULL, but I want to be safe.
 303          * This is a slow path anyway.
 304          */
 305         smp_wmb();
 306
 307         register_ftrace_function(&trace_ops);
 308
 309         if (tracing_is_enabled())
 310                 tracer_enabled = 1;
 311         else
 312                 tracer_enabled = 0;
 313
 314         return;
 315 fail_deprobe_wake_new:
 316         unregister_trace_sched_wakeup_new(probe_wakeup);
 317 fail_deprobe:
 318         unregister_trace_sched_wakeup(probe_wakeup);
 319 }
 320
 321 static void stop_wakeup_tracer(struct trace_array *tr)
 322 {
 323         tracer_enabled = 0;
 324         unregister_ftrace_function(&trace_ops);
 325         unregister_trace_sched_switch(probe_wakeup_sched_switch);
 326         unregister_trace_sched_wakeup_new(probe_wakeup);
 327         unregister_trace_sched_wakeup(probe_wakeup);
 328 }
 329
 330 static int __wakeup_tracer_init(struct trace_array *tr)
 331 {
 332         tracing_max_latency = 0;
 333         wakeup_trace = tr;
 334         start_wakeup_tracer(tr);
 335         return 0;
 336 }
 337
 338 static int wakeup_tracer_init(struct trace_array *tr)
 339 {
 340         wakeup_rt = 0;
 341         return __wakeup_tracer_init(tr);
 342 }
 343
 344 static int wakeup_rt_tracer_init(struct trace_array *tr)
 345 {
 346         wakeup_rt = 1;
 347         return __wakeup_tracer_init(tr);
 348 }
 349
 350 static void wakeup_tracer_reset(struct trace_array *tr)
 351 {
 352         stop_wakeup_tracer(tr);
 353         /* make sure we put back any tasks we are tracing */
 354         wakeup_reset(tr);
 355 }
 356
 357 static void wakeup_tracer_start(struct trace_array *tr)
 358 {
 359         wakeup_reset(tr);
 360         tracer_enabled = 1;
 361 }
 362
 363 static void wakeup_tracer_stop(struct trace_array *tr)
 364 {
 365         tracer_enabled = 0;
 366 }
 367
 368 static struct tracer wakeup_tracer __read_mostly =
 369 {
 370         .name           = "wakeup",
 371         .init           = wakeup_tracer_init,
 372         .reset          = wakeup_tracer_reset,
 373         .start          = wakeup_tracer_start,
 374         .stop           = wakeup_tracer_stop,
 375         .print_max      = 1,
 376 #ifdef CONFIG_FTRACE_SELFTEST
 377         .selftest    = trace_selftest_startup_wakeup,
 378 #endif
 379 };
 380
 381 static struct tracer wakeup_rt_tracer __read_mostly =
 382 {
 383         .name           = "wakeup_rt",
 384         .init           = wakeup_rt_tracer_init,
 385         .reset          = wakeup_tracer_reset,
 386         .start          = wakeup_tracer_start,
 387         .stop           = wakeup_tracer_stop,
 388         .print_max      = 1,
 389 #ifdef CONFIG_FTRACE_SELFTEST
 390         .selftest    = trace_selftest_startup_wakeup,
 391 #endif
 392 };
 393
 394 __init static int init_wakeup_tracer(void)
 395 {
 396         int ret;
 397
 398         ret = register_tracer(&wakeup_tracer);
 399         if (ret)
 400                 return ret;
 401
 402         ret = register_tracer(&wakeup_rt_tracer);
 403         if (ret)
 404                 return ret;
 405
 406         return 0;
 407 }
 408 device_initcall(init_wakeup_tracer);