git.oblomov.eu Git - 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 <linux/marker.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 wakeup_sched_switch(void *private, void *rq, struct task_struct *prev,
 116         struct task_struct *next)
 117 {
 118         unsigned long latency = 0, t0 = 0, t1 = 0;
 119         struct trace_array **ptr = private;
 120         struct trace_array *tr = *ptr;
 121         struct trace_array_cpu *data;
 122         cycle_t T0, T1, delta;
 123         unsigned long flags;
 124         long disabled;
 125         int cpu;
 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 = tr->data[wakeup_cpu];
 144
 145         /* disable local data, not wakeup_cpu data */
 146         cpu = raw_smp_processor_id();
 147         disabled = atomic_inc_return(&tr->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(tr, 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(tr, wakeup_task, wakeup_cpu);
 178
 179 out_unlock:
 180         __wakeup_reset(tr);
 181         __raw_spin_unlock(&wakeup_lock);
 182         local_irq_restore(flags);
 183 out:
 184         atomic_dec(&tr->data[cpu]->disabled);
 185 }
 186
 187 static notrace void
 188 sched_switch_callback(void *probe_data, void *call_data,
 189                       const char *format, va_list *args)
 190 {
 191         struct task_struct *prev;
 192         struct task_struct *next;
 193         struct rq *__rq;
 194
 195         /* skip prev_pid %d next_pid %d prev_state %ld */
 196         (void)va_arg(*args, int);
 197         (void)va_arg(*args, int);
 198         (void)va_arg(*args, long);
 199         __rq = va_arg(*args, typeof(__rq));
 200         prev = va_arg(*args, typeof(prev));
 201         next = va_arg(*args, typeof(next));
 202
 203         tracing_record_cmdline(prev);
 204
 205         /*
 206          * If tracer_switch_func only points to the local
 207          * switch func, it still needs the ptr passed to it.
 208          */
 209         wakeup_sched_switch(probe_data, __rq, prev, next);
 210 }
 211
 212 static void __wakeup_reset(struct trace_array *tr)
 213 {
 214         struct trace_array_cpu *data;
 215         int cpu;
 216
 217         for_each_possible_cpu(cpu) {
 218                 data = tr->data[cpu];
 219                 tracing_reset(data);
 220         }
 221
 222         wakeup_cpu = -1;
 223         wakeup_prio = -1;
 224
 225         if (wakeup_task)
 226                 put_task_struct(wakeup_task);
 227
 228         wakeup_task = NULL;
 229 }
 230
 231 static void wakeup_reset(struct trace_array *tr)
 232 {
 233         unsigned long flags;
 234
 235         local_irq_save(flags);
 236         __raw_spin_lock(&wakeup_lock);
 237         __wakeup_reset(tr);
 238         __raw_spin_unlock(&wakeup_lock);
 239         local_irq_restore(flags);
 240 }
 241
 242 static void
 243 wakeup_check_start(struct trace_array *tr, struct task_struct *p,
 244                    struct task_struct *curr)
 245 {
 246         int cpu = smp_processor_id();
 247         unsigned long flags;
 248         long disabled;
 249
 250         if (likely(!rt_task(p)) ||
 251                         p->prio >= wakeup_prio ||
 252                         p->prio >= curr->prio)
 253                 return;
 254
 255         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
 256         if (unlikely(disabled != 1))
 257                 goto out;
 258
 259         /* interrupts should be off from try_to_wake_up */
 260         __raw_spin_lock(&wakeup_lock);
 261
 262         /* check for races. */
 263         if (!tracer_enabled || p->prio >= wakeup_prio)
 264                 goto out_locked;
 265
 266         /* reset the trace */
 267         __wakeup_reset(tr);
 268
 269         wakeup_cpu = task_cpu(p);
 270         wakeup_prio = p->prio;
 271
 272         wakeup_task = p;
 273         get_task_struct(wakeup_task);
 274
 275         local_save_flags(flags);
 276
 277         tr->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
 278         trace_function(tr, tr->data[wakeup_cpu],
 279                        CALLER_ADDR1, CALLER_ADDR2, flags);
 280
 281 out_locked:
 282         __raw_spin_unlock(&wakeup_lock);
 283 out:
 284         atomic_dec(&tr->data[cpu]->disabled);
 285 }
 286
 287 static notrace void
 288 wake_up_callback(void *probe_data, void *call_data,
 289                  const char *format, va_list *args)
 290 {
 291         struct trace_array **ptr = probe_data;
 292         struct trace_array *tr = *ptr;
 293         struct task_struct *curr;
 294         struct task_struct *task;
 295         struct rq *__rq;
 296
 297         if (likely(!tracer_enabled))
 298                 return;
 299
 300         /* Skip pid %d state %ld */
 301         (void)va_arg(*args, int);
 302         (void)va_arg(*args, long);
 303         /* now get the meat: "rq %p task %p rq->curr %p" */
 304         __rq = va_arg(*args, typeof(__rq));
 305         task = va_arg(*args, typeof(task));
 306         curr = va_arg(*args, typeof(curr));
 307
 308         tracing_record_cmdline(task);
 309         tracing_record_cmdline(curr);
 310
 311         wakeup_check_start(tr, task, curr);
 312 }
 313
 314 static void start_wakeup_tracer(struct trace_array *tr)
 315 {
 316         int ret;
 317
 318         ret = marker_probe_register("kernel_sched_wakeup",
 319                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 320                         wake_up_callback,
 321                         &wakeup_trace);
 322         if (ret) {
 323                 pr_info("wakeup trace: Couldn't add marker"
 324                         " probe to kernel_sched_wakeup\n");
 325                 return;
 326         }
 327
 328         ret = marker_probe_register("kernel_sched_wakeup_new",
 329                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 330                         wake_up_callback,
 331                         &wakeup_trace);
 332         if (ret) {
 333                 pr_info("wakeup trace: Couldn't add marker"
 334                         " probe to kernel_sched_wakeup_new\n");
 335                 goto fail_deprobe;
 336         }
 337
 338         ret = marker_probe_register("kernel_sched_schedule",
 339                 "prev_pid %d next_pid %d prev_state %ld "
 340                 "## rq %p prev %p next %p",
 341                 sched_switch_callback,
 342                 &wakeup_trace);
 343         if (ret) {
 344                 pr_info("sched trace: Couldn't add marker"
 345                         " probe to kernel_sched_schedule\n");
 346                 goto fail_deprobe_wake_new;
 347         }
 348
 349         wakeup_reset(tr);
 350
 351         /*
 352          * Don't let the tracer_enabled = 1 show up before
 353          * the wakeup_task is reset. This may be overkill since
 354          * wakeup_reset does a spin_unlock after setting the
 355          * wakeup_task to NULL, but I want to be safe.
 356          * This is a slow path anyway.
 357          */
 358         smp_wmb();
 359
 360         register_ftrace_function(&trace_ops);
 361
 362         tracer_enabled = 1;
 363
 364         return;
 365 fail_deprobe_wake_new:
 366         marker_probe_unregister("kernel_sched_wakeup_new",
 367                                 wake_up_callback,
 368                                 &wakeup_trace);
 369 fail_deprobe:
 370         marker_probe_unregister("kernel_sched_wakeup",
 371                                 wake_up_callback,
 372                                 &wakeup_trace);
 373 }
 374
 375 static void stop_wakeup_tracer(struct trace_array *tr)
 376 {
 377         tracer_enabled = 0;
 378         unregister_ftrace_function(&trace_ops);
 379         marker_probe_unregister("kernel_sched_schedule",
 380                                 sched_switch_callback,
 381                                 &wakeup_trace);
 382         marker_probe_unregister("kernel_sched_wakeup_new",
 383                                 wake_up_callback,
 384                                 &wakeup_trace);
 385         marker_probe_unregister("kernel_sched_wakeup",
 386                                 wake_up_callback,
 387                                 &wakeup_trace);
 388 }
 389
 390 static void wakeup_tracer_init(struct trace_array *tr)
 391 {
 392         wakeup_trace = tr;
 393
 394         if (tr->ctrl)
 395                 start_wakeup_tracer(tr);
 396 }
 397
 398 static void wakeup_tracer_reset(struct trace_array *tr)
 399 {
 400         if (tr->ctrl) {
 401                 stop_wakeup_tracer(tr);
 402                 /* make sure we put back any tasks we are tracing */
 403                 wakeup_reset(tr);
 404         }
 405 }
 406
 407 static void wakeup_tracer_ctrl_update(struct trace_array *tr)
 408 {
 409         if (tr->ctrl)
 410                 start_wakeup_tracer(tr);
 411         else
 412                 stop_wakeup_tracer(tr);
 413 }
 414
 415 static void wakeup_tracer_open(struct trace_iterator *iter)
 416 {
 417         /* stop the trace while dumping */
 418         if (iter->tr->ctrl)
 419                 stop_wakeup_tracer(iter->tr);
 420 }
 421
 422 static void wakeup_tracer_close(struct trace_iterator *iter)
 423 {
 424         /* forget about any processes we were recording */
 425         if (iter->tr->ctrl)
 426                 start_wakeup_tracer(iter->tr);
 427 }
 428
 429 static struct tracer wakeup_tracer __read_mostly =
 430 {
 431         .name           = "wakeup",
 432         .init           = wakeup_tracer_init,
 433         .reset          = wakeup_tracer_reset,
 434         .open           = wakeup_tracer_open,
 435         .close          = wakeup_tracer_close,
 436         .ctrl_update    = wakeup_tracer_ctrl_update,
 437         .print_max      = 1,
 438 #ifdef CONFIG_FTRACE_SELFTEST
 439         .selftest    = trace_selftest_startup_wakeup,
 440 #endif
 441 };
 442
 443 __init static int init_wakeup_tracer(void)
 444 {
 445         int ret;
 446
 447         ret = register_tracer(&wakeup_tracer);
 448         if (ret)
 449                 return ret;
 450
 451         return 0;
 452 }
 453 device_initcall(init_wakeup_tracer);