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_FTRACE
  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
  48         if (likely(!wakeup_task))
  49                 return;
  50
  51         resched = need_resched();
  52         preempt_disable_notrace();
  53
  54         cpu = raw_smp_processor_id();
  55         data = tr->data[cpu];
  56         disabled = atomic_inc_return(&data->disabled);
  57         if (unlikely(disabled != 1))
  58                 goto out;
  59
  60         local_irq_save(flags);
  61         __raw_spin_lock(&wakeup_lock);
  62
  63         if (unlikely(!wakeup_task))
  64                 goto unlock;
  65
  66         /*
  67          * The task can't disappear because it needs to
  68          * wake up first, and we have the wakeup_lock.
  69          */
  70         if (task_cpu(wakeup_task) != cpu)
  71                 goto unlock;
  72
  73         trace_function(tr, data, ip, parent_ip, flags);
  74
  75  unlock:
  76         __raw_spin_unlock(&wakeup_lock);
  77         local_irq_restore(flags);
  78
  79  out:
  80         atomic_dec(&data->disabled);
  81
  82         /*
  83          * To prevent recursion from the scheduler, if the
  84          * resched flag was set before we entered, then
  85          * don't reschedule.
  86          */
  87         if (resched)
  88                 preempt_enable_no_resched_notrace();
  89         else
  90                 preempt_enable_notrace();
  91 }
  92
  93 static struct ftrace_ops trace_ops __read_mostly =
  94 {
  95         .func = wakeup_tracer_call,
  96 };
  97 #endif /* CONFIG_FTRACE */
  98
  99 /*
 100  * Should this new latency be reported/recorded?
 101  */
 102 static int report_latency(cycle_t delta)
 103 {
 104         if (tracing_thresh) {
 105                 if (delta < tracing_thresh)
 106                         return 0;
 107         } else {
 108                 if (delta <= tracing_max_latency)
 109                         return 0;
 110         }
 111         return 1;
 112 }
 113
 114 static void notrace
 115 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
 116         struct task_struct *next)
 117 {
 118         unsigned long latency = 0, t0 = 0, t1 = 0;
 119         struct trace_array_cpu *data;
 120         cycle_t T0, T1, delta;
 121         unsigned long flags;
 122         long disabled;
 123         int cpu;
 124
 125         tracing_record_cmdline(prev);
 126
 127         if (unlikely(!tracer_enabled))
 128                 return;
 129
 130         /*
 131          * When we start a new trace, we set wakeup_task to NULL
 132          * and then set tracer_enabled = 1. We want to make sure
 133          * that another CPU does not see the tracer_enabled = 1
 134          * and the wakeup_task with an older task, that might
 135          * actually be the same as next.
 136          */
 137         smp_rmb();
 138
 139         if (next != wakeup_task)
 140                 return;
 141
 142         /* The task we are waiting for is waking up */
 143         data = wakeup_trace->data[wakeup_cpu];
 144
 145         /* disable local data, not wakeup_cpu data */
 146         cpu = raw_smp_processor_id();
 147         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 148         if (likely(disabled != 1))
 149                 goto out;
 150
 151         local_irq_save(flags);
 152         __raw_spin_lock(&wakeup_lock);
 153
 154         /* We could race with grabbing wakeup_lock */
 155         if (unlikely(!tracer_enabled || next != wakeup_task))
 156                 goto out_unlock;
 157
 158         trace_function(wakeup_trace, data, CALLER_ADDR1, CALLER_ADDR2, flags);
 159
 160         /*
 161          * usecs conversion is slow so we try to delay the conversion
 162          * as long as possible:
 163          */
 164         T0 = data->preempt_timestamp;
 165         T1 = ftrace_now(cpu);
 166         delta = T1-T0;
 167
 168         if (!report_latency(delta))
 169                 goto out_unlock;
 170
 171         latency = nsecs_to_usecs(delta);
 172
 173         tracing_max_latency = delta;
 174         t0 = nsecs_to_usecs(T0);
 175         t1 = nsecs_to_usecs(T1);
 176
 177         update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
 178
 179 out_unlock:
 180         __wakeup_reset(wakeup_trace);
 181         __raw_spin_unlock(&wakeup_lock);
 182         local_irq_restore(flags);
 183 out:
 184         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 185 }
 186
 187 static void __wakeup_reset(struct trace_array *tr)
 188 {
 189         struct trace_array_cpu *data;
 190         int cpu;
 191
 192         for_each_possible_cpu(cpu) {
 193                 data = tr->data[cpu];
 194                 tracing_reset(tr, cpu);
 195         }
 196
 197         wakeup_cpu = -1;
 198         wakeup_prio = -1;
 199
 200         if (wakeup_task)
 201                 put_task_struct(wakeup_task);
 202
 203         wakeup_task = NULL;
 204 }
 205
 206 static void wakeup_reset(struct trace_array *tr)
 207 {
 208         unsigned long flags;
 209
 210         local_irq_save(flags);
 211         __raw_spin_lock(&wakeup_lock);
 212         __wakeup_reset(tr);
 213         __raw_spin_unlock(&wakeup_lock);
 214         local_irq_restore(flags);
 215 }
 216
 217 static void
 218 probe_wakeup(struct rq *rq, struct task_struct *p)
 219 {
 220         int cpu = smp_processor_id();
 221         unsigned long flags;
 222         long disabled;
 223
 224         if (likely(!tracer_enabled))
 225                 return;
 226
 227         tracing_record_cmdline(p);
 228         tracing_record_cmdline(current);
 229
 230         if (likely(!rt_task(p)) ||
 231                         p->prio >= wakeup_prio ||
 232                         p->prio >= current->prio)
 233                 return;
 234
 235         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
 236         if (unlikely(disabled != 1))
 237                 goto out;
 238
 239         /* interrupts should be off from try_to_wake_up */
 240         __raw_spin_lock(&wakeup_lock);
 241
 242         /* check for races. */
 243         if (!tracer_enabled || p->prio >= wakeup_prio)
 244                 goto out_locked;
 245
 246         /* reset the trace */
 247         __wakeup_reset(wakeup_trace);
 248
 249         wakeup_cpu = task_cpu(p);
 250         wakeup_prio = p->prio;
 251
 252         wakeup_task = p;
 253         get_task_struct(wakeup_task);
 254
 255         local_save_flags(flags);
 256
 257         wakeup_trace->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
 258         trace_function(wakeup_trace, wakeup_trace->data[wakeup_cpu],
 259                        CALLER_ADDR1, CALLER_ADDR2, flags);
 260
 261 out_locked:
 262         __raw_spin_unlock(&wakeup_lock);
 263 out:
 264         atomic_dec(&wakeup_trace->data[cpu]->disabled);
 265 }
 266
 267 static void start_wakeup_tracer(struct trace_array *tr)
 268 {
 269         int ret;
 270
 271         ret = register_trace_sched_wakeup(probe_wakeup);
 272         if (ret) {
 273                 pr_info("wakeup trace: Couldn't activate tracepoint"
 274                         " probe to kernel_sched_wakeup\n");
 275                 return;
 276         }
 277
 278         ret = register_trace_sched_wakeup_new(probe_wakeup);
 279         if (ret) {
 280                 pr_info("wakeup trace: Couldn't activate tracepoint"
 281                         " probe to kernel_sched_wakeup_new\n");
 282                 goto fail_deprobe;
 283         }
 284
 285         ret = register_trace_sched_switch(probe_wakeup_sched_switch);
 286         if (ret) {
 287                 pr_info("sched trace: Couldn't activate tracepoint"
 288                         " probe to kernel_sched_schedule\n");
 289                 goto fail_deprobe_wake_new;
 290         }
 291
 292         wakeup_reset(tr);
 293
 294         /*
 295          * Don't let the tracer_enabled = 1 show up before
 296          * the wakeup_task is reset. This may be overkill since
 297          * wakeup_reset does a spin_unlock after setting the
 298          * wakeup_task to NULL, but I want to be safe.
 299          * This is a slow path anyway.
 300          */
 301         smp_wmb();
 302
 303         register_ftrace_function(&trace_ops);
 304
 305         tracer_enabled = 1;
 306
 307         return;
 308 fail_deprobe_wake_new:
 309         unregister_trace_sched_wakeup_new(probe_wakeup);
 310 fail_deprobe:
 311         unregister_trace_sched_wakeup(probe_wakeup);
 312 }
 313
 314 static void stop_wakeup_tracer(struct trace_array *tr)
 315 {
 316         tracer_enabled = 0;
 317         unregister_ftrace_function(&trace_ops);
 318         unregister_trace_sched_switch(probe_wakeup_sched_switch);
 319         unregister_trace_sched_wakeup_new(probe_wakeup);
 320         unregister_trace_sched_wakeup(probe_wakeup);
 321 }
 322
 323 static void wakeup_tracer_init(struct trace_array *tr)
 324 {
 325         wakeup_trace = tr;
 326
 327         if (tr->ctrl)
 328                 start_wakeup_tracer(tr);
 329 }
 330
 331 static void wakeup_tracer_reset(struct trace_array *tr)
 332 {
 333         if (tr->ctrl) {
 334                 stop_wakeup_tracer(tr);
 335                 /* make sure we put back any tasks we are tracing */
 336                 wakeup_reset(tr);
 337         }
 338 }
 339
 340 static void wakeup_tracer_ctrl_update(struct trace_array *tr)
 341 {
 342         if (tr->ctrl)
 343                 start_wakeup_tracer(tr);
 344         else
 345                 stop_wakeup_tracer(tr);
 346 }
 347
 348 static void wakeup_tracer_open(struct trace_iterator *iter)
 349 {
 350         /* stop the trace while dumping */
 351         if (iter->tr->ctrl)
 352                 stop_wakeup_tracer(iter->tr);
 353 }
 354
 355 static void wakeup_tracer_close(struct trace_iterator *iter)
 356 {
 357         /* forget about any processes we were recording */
 358         if (iter->tr->ctrl)
 359                 start_wakeup_tracer(iter->tr);
 360 }
 361
 362 static struct tracer wakeup_tracer __read_mostly =
 363 {
 364         .name           = "wakeup",
 365         .init           = wakeup_tracer_init,
 366         .reset          = wakeup_tracer_reset,
 367         .open           = wakeup_tracer_open,
 368         .close          = wakeup_tracer_close,
 369         .ctrl_update    = wakeup_tracer_ctrl_update,
 370         .print_max      = 1,
 371 #ifdef CONFIG_FTRACE_SELFTEST
 372         .selftest    = trace_selftest_startup_wakeup,
 373 #endif
 374 };
 375
 376 __init static int init_wakeup_tracer(void)
 377 {
 378         int ret;
 379
 380         ret = register_tracer(&wakeup_tracer);
 381         if (ret)
 382                 return ret;
 383
 384         return 0;
 385 }
 386 device_initcall(init_wakeup_tracer);