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
  29 static raw_spinlock_t wakeup_lock =
  30         (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  31
  32 static void __wakeup_reset(struct trace_array *tr);
  33
  34 #ifdef CONFIG_FUNCTION_TRACER
  35 /*
  36  * irqsoff uses its own tracer function to keep the overhead down:
  37  */
  38 static void
  39 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
  40 {
  41         struct trace_array *tr = wakeup_trace;
  42         struct trace_array_cpu *data;
  43         unsigned long flags;
  44         long disabled;
  45         int resched;
  46         int cpu;
  47         int pc;
  48
  49         if (likely(!wakeup_task))
  50                 return;
  51
  52         pc = preempt_count();
  53         resched = ftrace_preempt_disable();
  54
  55         cpu = raw_smp_processor_id();
  56         data = tr->data[cpu];
  57         disabled = atomic_inc_return(&data->disabled);
  58         if (unlikely(disabled != 1))
  59                 goto out;
  60
  61         local_irq_save(flags);
  62         __raw_spin_lock(&wakeup_lock);
  63
  64         if (unlikely(!wakeup_task))
  65                 goto unlock;
  66
  67         /*
  68          * The task can't disappear because it needs to
  69          * wake up first, and we have the wakeup_lock.
  70          */
  71         if (task_cpu(wakeup_task) != cpu)
  72                 goto unlock;
  73
  74         trace_function(tr, data, ip, parent_ip, flags, pc);
  75
  76  unlock:
  77         __raw_spin_unlock(&wakeup_lock);
  78         local_irq_restore(flags);
  79
  80  out:
  81         atomic_dec(&data->disabled);
  82
  83         ftrace_preempt_enable(resched);
  84 }
  85
  86 static struct ftrace_ops trace_ops __read_mostly =
  87 {
  88         .func = wakeup_tracer_call,
  89 };
  90 #endif /* CONFIG_FUNCTION_TRACER */
  91
  92 /*
  93  * Should this new latency be reported/recorded?
  94  */
  95 static int report_latency(cycle_t delta)
  96 {
  97         if (tracing_thresh) {
  98                 if (delta < tracing_thresh)
  99                         return 0;
 100         } else {
 101                 if (delta <= tracing_max_latency)
 102                         return 0;
 103         }
 104         return 1;
 105 }
 106
 107 static void notrace
 108 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
 109         struct task_struct *next)
 110 {
 111         unsigned long latency = 0, t0 = 0, t1 = 0;
 112         struct trace_array_cpu *data;
 113         cycle_t T0, T1, delta;
 114         unsigned long flags;
 115         long disabled;
 116         int cpu;
 117         int pc;
 118
 119         tracing_record_cmdline(prev);
 120
 121         if (unlikely(!tracer_enabled))
 122                 return;
 123
 124         /*
 125          * When we start a new trace, we set wakeup_task to NULL
 126          * and then set tracer_enabled = 1. We want to make sure
 127          * that another CPU does not see the tracer_enabled = 1
 128          * and the wakeup_task with an older task, that might
 129          * actually be the same as next.
 130          */
 131         smp_rmb();
 132
 133         if (next != wakeup_task)
 134                 return;
 135
 136         pc = preempt_count();
 137
 138         /* The task we are waiting for is waking up */
 139         data = wakeup_trace->data[wakeup_cpu];
 140
 141         /* disable local data, not wakeup_cpu data */
 142         cpu = raw_smp_processor_id();
 143         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 144         if (likely(disabled != 1))
 145                 goto out;
 146
 147         local_irq_save(flags);
 148         __raw_spin_lock(&wakeup_lock);
 149
 150         /* We could race with grabbing wakeup_lock */
 151         if (unlikely(!tracer_enabled || next != wakeup_task))
 152                 goto out_unlock;
 153
 154         trace_function(wakeup_trace, data, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
 155
 156         /*
 157          * usecs conversion is slow so we try to delay the conversion
 158          * as long as possible:
 159          */
 160         T0 = data->preempt_timestamp;
 161         T1 = ftrace_now(cpu);
 162         delta = T1-T0;
 163
 164         if (!report_latency(delta))
 165                 goto out_unlock;
 166
 167         latency = nsecs_to_usecs(delta);
 168
 169         tracing_max_latency = delta;
 170         t0 = nsecs_to_usecs(T0);
 171         t1 = nsecs_to_usecs(T1);
 172
 173         update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
 174
 175 out_unlock:
 176         __wakeup_reset(wakeup_trace);
 177         __raw_spin_unlock(&wakeup_lock);
 178         local_irq_restore(flags);
 179 out:
 180         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 181 }
 182
 183 static void __wakeup_reset(struct trace_array *tr)
 184 {
 185         struct trace_array_cpu *data;
 186         int cpu;
 187
 188         for_each_possible_cpu(cpu) {
 189                 data = tr->data[cpu];
 190                 tracing_reset(tr, cpu);
 191         }
 192
 193         wakeup_cpu = -1;
 194         wakeup_prio = -1;
 195
 196         if (wakeup_task)
 197                 put_task_struct(wakeup_task);
 198
 199         wakeup_task = NULL;
 200 }
 201
 202 static void wakeup_reset(struct trace_array *tr)
 203 {
 204         unsigned long flags;
 205
 206         local_irq_save(flags);
 207         __raw_spin_lock(&wakeup_lock);
 208         __wakeup_reset(tr);
 209         __raw_spin_unlock(&wakeup_lock);
 210         local_irq_restore(flags);
 211 }
 212
 213 static void
 214 probe_wakeup(struct rq *rq, struct task_struct *p, int success)
 215 {
 216         int cpu = smp_processor_id();
 217         unsigned long flags;
 218         long disabled;
 219         int pc;
 220
 221         if (likely(!tracer_enabled))
 222                 return;
 223
 224         tracing_record_cmdline(p);
 225         tracing_record_cmdline(current);
 226
 227         if (likely(!rt_task(p)) ||
 228                         p->prio >= wakeup_prio ||
 229                         p->prio >= current->prio)
 230                 return;
 231
 232         pc = preempt_count();
 233         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 234         if (unlikely(disabled != 1))
 235                 goto out;
 236
 237         /* interrupts should be off from try_to_wake_up */
 238         __raw_spin_lock(&wakeup_lock);
 239
 240         /* check for races. */
 241         if (!tracer_enabled || p->prio >= wakeup_prio)
 242                 goto out_locked;
 243
 244         /* reset the trace */
 245         __wakeup_reset(wakeup_trace);
 246
 247         wakeup_cpu = task_cpu(p);
 248         wakeup_prio = p->prio;
 249
 250         wakeup_task = p;
 251         get_task_struct(wakeup_task);
 252
 253         local_save_flags(flags);
 254
 255         wakeup_trace->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
 256         trace_function(wakeup_trace, wakeup_trace->data[wakeup_cpu],
 257                        CALLER_ADDR1, CALLER_ADDR2, flags, pc);
 258
 259 out_locked:
 260         __raw_spin_unlock(&wakeup_lock);
 261 out:
 262         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 263 }
 264
 265 /*
 266  * save_tracer_enabled is used to save the state of the tracer_enabled
 267  * variable when we disable it when we open a trace output file.
 268  */
 269 static int save_tracer_enabled;
 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                 save_tracer_enabled = 1;
 312         } else {
 313                 tracer_enabled = 0;
 314                 save_tracer_enabled = 0;
 315         }
 316
 317         return;
 318 fail_deprobe_wake_new:
 319         unregister_trace_sched_wakeup_new(probe_wakeup);
 320 fail_deprobe:
 321         unregister_trace_sched_wakeup(probe_wakeup);
 322 }
 323
 324 static void stop_wakeup_tracer(struct trace_array *tr)
 325 {
 326         tracer_enabled = 0;
 327         save_tracer_enabled = 0;
 328         unregister_ftrace_function(&trace_ops);
 329         unregister_trace_sched_switch(probe_wakeup_sched_switch);
 330         unregister_trace_sched_wakeup_new(probe_wakeup);
 331         unregister_trace_sched_wakeup(probe_wakeup);
 332 }
 333
 334 static int wakeup_tracer_init(struct trace_array *tr)
 335 {
 336         tracing_max_latency = 0;
 337         wakeup_trace = tr;
 338         start_wakeup_tracer(tr);
 339         return 0;
 340 }
 341
 342 static void wakeup_tracer_reset(struct trace_array *tr)
 343 {
 344         stop_wakeup_tracer(tr);
 345         /* make sure we put back any tasks we are tracing */
 346         wakeup_reset(tr);
 347 }
 348
 349 static void wakeup_tracer_start(struct trace_array *tr)
 350 {
 351         wakeup_reset(tr);
 352         tracer_enabled = 1;
 353         save_tracer_enabled = 1;
 354 }
 355
 356 static void wakeup_tracer_stop(struct trace_array *tr)
 357 {
 358         tracer_enabled = 0;
 359         save_tracer_enabled = 0;
 360 }
 361
 362 static void wakeup_tracer_open(struct trace_iterator *iter)
 363 {
 364         /* stop the trace while dumping */
 365         tracer_enabled = 0;
 366 }
 367
 368 static void wakeup_tracer_close(struct trace_iterator *iter)
 369 {
 370         /* forget about any processes we were recording */
 371         if (save_tracer_enabled) {
 372                 wakeup_reset(iter->tr);
 373                 tracer_enabled = 1;
 374         }
 375 }
 376
 377 static struct tracer wakeup_tracer __read_mostly =
 378 {
 379         .name           = "wakeup",
 380         .init           = wakeup_tracer_init,
 381         .reset          = wakeup_tracer_reset,
 382         .start          = wakeup_tracer_start,
 383         .stop           = wakeup_tracer_stop,
 384         .open           = wakeup_tracer_open,
 385         .close          = wakeup_tracer_close,
 386         .print_max      = 1,
 387 #ifdef CONFIG_FTRACE_SELFTEST
 388         .selftest    = trace_selftest_startup_wakeup,
 389 #endif
 390 };
 391
 392 __init static int init_wakeup_tracer(void)
 393 {
 394         int ret;
 395
 396         ret = register_tracer(&wakeup_tracer);
 397         if (ret)
 398                 return ret;
 399
 400         return 0;
 401 }
 402 device_initcall(init_wakeup_tracer);